Refactor GeoTIFF Phase 5c-write: extract _write_layout.py from _writer.py

[brendancol]

Closes #2248
Part of #2211

Summary

• Moves IFD encoding primitives and BigTIFF / COG layout planners out of xrspatial/geotiff/_writer.py into a new xrspatial/geotiff/_write_layout.py.
• Pixel encoding kernels (strip/tile compression, photometric/predictor encode) and the top-level _write / _write_streaming entry points stay in _writer.py.
• _writer.py drops about 580 lines.

Moved names

_float_to_rational, _serialize_tag_value, _pack_tag_value, _build_ifd, _assemble_tiff, _promote_offsets_to_long8, _assemble_standard_layout, _assemble_cog_layout, _compute_classic_ifd_overhead, _should_use_bigtiff_streaming, plus the BO byte-order constant and the _BIGTIFF_OFFSET_TAGS frozenset.

Compatibility

• _writer.py re-exports every moved name. Existing imports (_writers/eager.py, _writers/gpu.py, _gpu_decode.py, and the test suite that imports from xrspatial.geotiff._writer) keep working unchanged.
• _assemble_tiff resolves the layout helpers and _resolve_photometric / _OVERRIDABLE_AUTO_TAG_IDS / _DANGEROUS_EXTRA_TAG_IDS through the _writer module at call time so tests that monkeypatch _writer.* (e.g. test_eager_bigtiff_overhead_exact_1905.py) keep their pre-extraction semantics.
• No public API change.

Test plan

• pytest xrspatial/geotiff/tests/ passes locally (5033 passed, 68 skipped). One pre-existing failure unrelated to this change (test_lowlevel_write_pushdown_2138.py::test_write_vs_to_geotiff_byte_parity_uint8[lz4], reproduces on main) was deselected.
• Module imports cleanly: from xrspatial.geotiff._writer import _assemble_tiff, _build_ifd, _compute_classic_ifd_overhead still resolves to the moved implementations.
• CI green.

[brendancol]

PR Review: Extract _write_layout.py from _writer.py

Blockers

None.

Suggestions

1. _write_layout.py:497-504 - _assemble_tiff reaches back into _writer for six names every call so test monkeypatches on _writer.* still take effect. The contract is invisible unless you read the comment block, and it couples the new module back to the one it was extracted from. Two options: (a) call this out on the module docstring so nobody inlines _compute_classic_ifd_overhead here in a future cleanup, or (b) add a unit test that monkeypatches one of the other indirected names (e.g. _writer._promote_offsets_to_long8 or _writer._assemble_standard_layout) and asserts the override fires through _assemble_tiff. Right now only the _compute_classic_ifd_overhead path is exercised by test_eager_bigtiff_overhead_exact_1905.

2. _write_layout.py:194-200 - _BIGTIFF_OFFSET_TAGS is consumed only by _promote_offsets_to_long8 and has no external importers. The # noqa: F401 re-export is harmless, but the constant could come off the re-export list; the public surface is _promote_offsets_to_long8. Not blocking.

Nits

1. _writer.py:42-50 - Dropped LONG8 from the _dtypes import block but left TIFF_TYPE_SIZES. It was already unused on main, so this PR didn't introduce the dead import, but cleaning it up here would save the next reader a double-take.

2. _write_layout.py:1-12 - Module docstring says the helpers feed "the eager and streaming writers." True, but _write_streaming calls the lower-level helpers (_build_ifd, _compute_classic_ifd_overhead, _should_use_bigtiff_streaming, _promote_offsets_to_long8) directly and never goes through _assemble_tiff. A one-line note that "the eager writer calls _assemble_tiff; the streaming writer composes the lower-level helpers" would make the split clearer.

3. _write_layout.py:497 - from . import _writer as _writer_mod works but is unusual. A short docstring sentence on _assemble_tiff (something like "this function is owned by _writer.py semantically; it lives here only to keep IFD-encoding helpers co-located") would help reviewers who land in the file directly.

What looks good

• Mechanical and behaviour-neutral. git diff --stat reports 611 deletions and 29 insertions in _writer.py, inside the issue's 600-900-line target.
• Re-exports cover every existing import path: _writers/eager.py, _writers/gpu.py, _gpu_decode.py, plus the test imports in tests/test_features.py, tests/test_assemble_layout_no_bytes_copy_1756.py, tests/test_eager_bigtiff_overhead_exact_1905.py, tests/test_predictor3_int_dtype_1933.py, tests/test_predictor3_int_dtype_gpu_1933.py.
• The monkeypatch-preservation indirection in _assemble_tiff was the pragmatic option; the alternative was patching every test call site, which would balloon the PR.
• 5033 geotiff tests pass. The single pre-existing lz4 failure also reproduces on main and is unrelated.

Checklist

• No public API change.
• Re-exports preserve every existing import path.
• Eager and streaming write paths covered by the geotiff suite.
• Monkeypatch contract on _writer.* preserved.
• No new public functions, so the README feature matrix is unaffected.
• No new docs entries needed (private module).

[brendancol]

PR Review (follow-up): #2249

Blockers

None.

Suggestions

None.

Nits

1. tests/test_write_layout_monkeypatch_contract_2248.py:60-62 - The _wrapped sentinel records args and kwargs but the calls list is only used for a truthiness check. Could drop the recording and use a simple nonlocal called = True flag. Not blocking; the captured arguments may help if the test ever needs to assert against the call shape.

What looks good (delta from first pass)

• All actionable items from the first review applied:
  • _BIGTIFF_OFFSET_TAGS removed from the _writer.py re-export (_writer.py:89-101).
  • TIFF_TYPE_SIZES dead import dropped (_writer.py:42-49).
  • _write_layout.py module docstring now spells out eager vs. streaming consumers and documents the _writer-module indirection contract.
  • _assemble_tiff docstring explains the ownership / co-location rationale.
• New regression test (tests/test_write_layout_monkeypatch_contract_2248.py) parametrises across the four previously-uncovered indirected names (_promote_offsets_to_long8, _assemble_standard_layout, _assemble_cog_layout, _resolve_photometric). It picks the right kwargs per helper (cog=True, overview_levels=[2] for COG, bigtiff=True for the BigTIFF-only promoter) so each helper actually fires on its subtest. With the existing test_eager_bigtiff_overhead_exact_1905, all five indirected names are locked down.
• 5037 tests pass locally. The lz4 pre-existing failure on test_lowlevel_write_pushdown_2138.py still reproduces on main and is unrelated.

Checklist

• No public API change.
• Re-exports preserve every existing import path.
• Eager and streaming write paths covered by the geotiff suite.
• Monkeypatch contract on _writer.* covered by tests for all five indirected names.
• No new public functions, so the README feature matrix is unaffected.
• No new docs entries needed (private module).

Verdict: clean / nits-only. The single nit is cosmetic; not worth another round.