[None][feat] add commit-min snapshots and per-conversation KV reuse

[jiaganc]

@coderabbitai summary

Description

Integrates commit_min_snapshot support from #15752 and per-conversation KV cache block reuse from #16115 into the DeepSeek-V4 feature branch.

The runtime now supports minimum snapshots for SSM reuse, including partial final-page commits. It also exposes CommittedBlockRecord, which keeps selected committed SWA-window pages reusable until the record is released. Full-attention and attention-sink pages remain managed by the existing global reuse policy; recording SSM state for a conversation remains a TODO.

The PyTorch KV cache manager V2 now:

• propagates partial-reuse configuration to the generic and DeepSeek-V4 cache managers;
• enables commit_min_snapshot when block reuse is enabled and the policy is not all_reusable;
• supports all_reusable, per_request, and per_conversation policies;
• delays per_request and per_conversation commits until the final context chunk while allowing a partial final commit;
• tracks the active request and latest committed-block record for each conversation;
• replaces the previous conversation record only after a completed turn is recorded successfully;
• ignores conversation tracking for overlapping requests in the same conversation and uses per_request behavior when conversation parameters are absent.

The executor also documents a follow-up optimization: a disaggregated prefill server can pass is_end=True for the final context chunk to avoid unnecessary final-page copies.

Test Coverage

• Added minimum-snapshot and partial-reuse configuration coverage for generic V2 and DeepSeek-V4 cache managers.
• Added partial final-context commit and reuse-policy timing coverage.
• Added committed-block record lifecycle coverage in tests/unittest/kv_cache_manager_v2_tests/test_kv_cache_manager_v2.py.
• Added per-conversation chunking, fallback, divergent-turn, and overlapping-request coverage in tests/unittest/_torch/executor/test_kv_cache_manager_v2.py.
• Extended KvCacheConfig validation coverage for per_conversation.
• Python compilation, git diff --check, and all relevant pre-commit hooks passed locally.
• Pytest was not run locally because the checkout's Python environment does not provide the pytest module.

PR Checklist

Please review the following before submitting this PR:

• PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.

• PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

• Test cases are provided for new code paths (see test instructions).

• Any new dependencies have been scanned for license and vulnerabilities.

• CODEOWNERS updated if ownership changes.

• Documentation updated as needed.

• Update tava architecture diagram if there is a significant design change in PR.

• The reviewers assigned automatically/manually are appropriate for the PR.

• Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.

[longlee0622]

We no longer use the ds v4 side branch. Closing this PR. Please prepare PR against main.