[codex] fix(input): buffer Windows console decoder bytes

[tonyfettes]

What changed

• Add internal/io.ByteQueue, an internal byte source that implements @async/io.Reader and supports synchronous byte enqueueing.
• Replace the Windows console input byte pipe with ByteQueue so key-record bytes are buffered before the shared input decoder consumes them.
• Drain available Windows console records synchronously and queue direct native events with try_put.
• Add regression coverage for queued direct events followed by a complete SGR mouse byte sequence.
• Update examples/raw to show printable ASCII characters next to raw byte values.

Why

Windows ReadConsoleInputW key records can represent VT mouse sequences as individual UTF-16 characters. The previous unbuffered pipe write could suspend inside the short-lived record producer; if a direct event won the read race, the producer could be cancelled before the tail of the VT sequence reached the decoder. That made examples/input display fragments such as [3;20M as key events.

Buffering bytes in an internal queue makes producer-side enqueueing synchronous, while preserving the existing shared input decoder and direct native event queue.

Validation

• moon fmt
• moon test internal/io --target-dir .moon-test-internal-io-bytequeue-build
• moon test . --filter "win32*" --target-dir .moon-test-win32-bytequeue-shape-build
• moon test internal/input --filter "decode delayed *mouse*" --target-dir .moon-test-input-delayed-mouse-bytequeue-shape-build
• moon check --target-dir .moon-check-bytequeue-shape-build
• moon info --target-dir .moon-info-bytequeue-shape-build
• git diff --check

[tonyfettes]

1. Don't eats up error without testing for @async.is_being_cancelled.
2. Use if ... is, don't use match on Option
3. Use guard to de-indent code.

[tonyfettes]

Addressed the latest review feedback in 15a7926, corrected the ByteQueue visibility in 74b6916, and removed the unnecessary cancellation shield in a267bf8.

Notes on the non-literal bits:

• ByteQueue is now a plain struct, which generates type ByteQueue in internal/io/pkg.generated.mbti. That keeps it usable by the root package through its public methods while hiding its representation.
• ByteQueue._direct_read relies on Queue::get() for the wake/cancel race, then performs the buffer write and drain synchronously before returning len.
• Removed the cancellable producer task from the Win32 decoder tests by queueing records synchronously before reading.
• Replaced the unsafe char conversions with a checked Int::to_char() helper.

[tonyfettes]

Why we protected from cancel here?