[None][feat] Add V2 Mamba hybrid cache with snapshot reuse

[VALLIS-NERIA]

@coderabbitai summary

Description

This PR adds the V2 Mamba hybrid-cache core and snapshot reuse on top of the generic scheduler/resource-manager contracts in #16427.

The changes:

• Add V2MambaHybridCacheManager, which stores attention KV blocks together with Mamba convolution and SSM state in KVCMv2-managed pools.
• Expose role-aware pool addresses and block indices so linear-attention state can use logical views over the underlying block layout.
• Track live state and reusable snapshots separately, commit snapshots only at reusable boundaries, and advance snapshot history consistently across decode and MTP replay.
• Keep CUDA state-index storage stable across attention-DP dummy insertion and remove request-to-index mappings whenever requests are freed.
• Preserve explicit C++/mixed-manager routing and legacy disaggregated-serving fallbacks; this PR does not add V2 Mamba state transfer.
• Add focused coverage for allocation, snapshot lookup/commit, replay state, dummy indices, scheduler chunk boundaries, stats, and transceiver integration.

This PR depends on #16427. GitHub cannot use a branch from a fork as the base of an upstream PR, so the draft diff temporarily includes #16427 until that PR merges and this branch is rebased. The commit-minimum integration formerly carried by this work is already available on main through #16185.

Save-last snapshot reuse and V2 Mamba disaggregated transfer are intentionally split into follow-up PRs.

Test Coverage

• pytest -q tests/unittest/_torch/executor/test_mamba_cache_manager.py (44 passed)
• pytest -q tests/unittest/_torch/executor/test_kv_cache_manager_v2.py (10 passed)
• pytest -q tests/unittest/kv_cache_manager_v2_tests/test_kv_cache_stats_behavior.py (18 passed)
• pytest -q tests/unittest/kv_cache_manager_v2_tests/test_kv_cache_manager_v2.py (102 passed; the one stale-config failure was fixed and its focused rerun passed)
• Focused Python/NIXL native transceiver tests after rebuilding bindings against current main (7 passed)
• pre-commit run --from-ref origin/main --to-ref HEAD

PR Checklist

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

[VALLIS-NERIA]

/bot run

[tensorrt-cicd]

PR_Github #59619 [ run ] triggered by Bot. Commit: 1461b2d Link to invocation

[tensorrt-cicd]

PR_Github #59619 [ run ] completed with state SUCCESS. Commit: 1461b2d
/LLM/main/L0_MergeRequest_PR pipeline #48056 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

[VALLIS-NERIA]

/bot run

[tensorrt-cicd]

PR_Github #59668 [ run ] triggered by Bot. Commit: 149013b Link to invocation

[VALLIS-NERIA]

Reopened as a stacked draft at VALLIS-NERIA#8 so it can target the PR1 branch and show only its single squashed commit. Closing this upstream draft as superseded.

[tensorrt-cicd]

PR_Github #59668 [ run ] completed with state SUCCESS. Commit: 149013b
/LLM/main/L0_MergeRequest_PR pipeline #48101 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation