feat(loss): add pg_loss aggregation modes

[EazyReal]

On main, slime has two built-in pg_loss normalizations: the default sample/rollout mean and the legacy --calculate-per-token-loss path. Prompt-group normalization and fixed-divisor normalization both require a custom pg_loss reducer today. That is fragile because the reducer has to compose correctly with CP slicing, micro-batch packing, and Megatron's final train-step divisor; a wrong constant factor silently changes the effective learning rate.

This PR moves the common pg_loss aggregation choices behind --loss-aggregation. For one train step, let N be the number of rollouts, G = n_samples_per_prompt, P = N / G, M_i be the valid-token count for rollout i, and L = --loss-aggregation-divisor.

mode	step scalar	behavior
sample_mean	(1 / N) * sum_i token_mean(i)	Keeps the current default behavior.
token_mean	sum_i,t loss_it * mask_it / sum_i M_i	Uses the existing --calculate-per-token-loss path under the new unified knob.
prompt_mean	(1 / P) * sum_g token_mean(prompt_group_g)	Groups rollouts by Sample.group_index and gives each prompt group one unit of weight.
constant	sum_i,t loss_it * mask_it / (L * N)	Uses a fixed token-scale divisor before the usual step average.

The implementation reuses slime's existing reducer path instead of adding a parallel loss stack. Rollout builds prompt_mask_sums only for prompt_mean; cp_utils.get_sum_of_sample_mean still owns CP-aware summation; and the train step keeps its existing final scaling. For prompt_mean, the reducer scales the prompt-group sum by n_samples_per_prompt, so users get the requested prompt mean directly instead of a constant-offset objective that would need learning-rate compensation.

Startup validation rejects configurations that would silently use the wrong denominator, such as --loss-aggregation-divisor outside constant or --calculate-per-token-loss together with prompt_mean/constant. The custom pg_loss reducer hook remains available and keeps precedence for non-standard objectives.

The focused tests cover the final prompt_mean scalar, CP/sample-denominator invariance, argument validation, and rollout metadata construction.

Tested with:

• uv run --no-project --with pytest --with torch --with numpy --with httpx --with pyyaml python -m pytest -q tests/test_cp_utils.py tests/test_megatron_argument_validation.py tests/test_rollout_validation.py
• uv run --no-project --with ruff ruff check --config pyproject.toml docs/en/get_started/customization.md slime/backends/megatron_utils/loss.py slime/ray/rollout.py slime/utils/arguments.py tests/test_cp_utils.py tests/test_megatron_argument_validation.py tests/test_rollout_validation.py
• uv run --no-project --with black black --check --line-length 119 slime/backends/megatron_utils/loss.py slime/utils/arguments.py