feat(linear-logp): add strict bitwise SM90 contract

[inaniloquentee]

Summary

• freeze the SM90 fused linear_logp reduction contract so vocab splitting and reduction order no longer depend on batch shape, occupancy, or SM count
• apply temperature at the same FP32 epilogue point, optionally return sampler logits, and mask padded vocabulary lanes out of the LSE
• expose separate strict single-rank and tensor-parallel entry points while preserving the existing fast path
• make TP LSE and selected-logit merges explicitly rank ordered and keep strict autograd mathematically correct

Refs #335

Validation

• built the CUDA extension from source for CUDA 12.4 / SM90 with KERNEL_ALIGN_FORCE_SM90=1
• python -m pytest tests/test_linear_logp.py -q (41 passed)
• strict tests compare raw FP32 bit patterns across full-batch and single-row execution, return_logits modes, temperature, and padded-vocab masking
• flake8 on changed Python files and git diff --check

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: cf87c042-daa0-4df1-90f3-a3e8d918ca52

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[Flink-ddd]

LGTM

[maxiaosong1124]

LGTM!