Fix partial rollout off-policy loss mask alignment

[chengcuiping]

Summary

• Add a small Sample.mask_response_tokens() helper that preserves response-side loss_mask length invariants.
• Use it when partial rollout masks previous off-policy response tokens.
• Add CPU regression coverage for partial resume appending new trainable tokens after masking old tokens.

Motivation

When partial rollout is combined with off-policy masking, old generated tokens are masked to zero before resumed generation appends new response tokens. If response_length grows without loss_mask staying aligned, training conversion can fail with len(sample.loss_mask) == sample.response_length, as reported in #1440.

This is a small stability step for the partial rollout / off-policy trajectory work discussed in #1800.

Behavior

• Old partial response tokens remain masked to 0.
• Newly resumed response tokens append trainable 1s through append_response_tokens().
• len(loss_mask) == response_length is preserved.
• Non-partial rollout behavior is unchanged.
• No RL algorithm, KV cache, routing, or schema changes.

Testing

• CUDA_VISIBLE_DEVICES='' python -m pytest tests/test_partial_rollout_loss_mask.py tests/test_sample.py -m unit
• CUDA_VISIBLE_DEVICES='' pre-commit run --files slime/utils/types.py slime/rollout/sglang_rollout.py tests/test_partial_rollout_loss_mask.py
• git diff --check

All passed.

tests/test_rollout_metrics.py could not be collected in this environment because the installed sglang package is missing GPU_MEMORY_TYPE_CUDA_GRAPH from sglang.srt.constants.

Refs #1440, #1800.