[https://nvbugs/6438658][fix] Fix KV cache estimation capacity

[jiaganc]

@coderabbitai summary

Description

Fix KV cache capacity estimation so the resulting cache can satisfy runtime warmup and resume requirements.

Without these constraints, pool-ratio estimation can allocate a pool layout that cannot run CUDA graph warmup. Ignoring resume utilization during distributed quota synchronization can also remove the reserved headroom and overestimate usable token capacity.

• Preserve max_util_for_resume headroom when synchronizing KV cache capacity across distributed ranks.
• Create the generic V2 cache configuration, standard layers, and estimation constraints in KVCacheManagerV2.
• Cover the full CUDA graph generation warmup batch and the per-iteration context token budget when estimating pool ratios.
• Let specialized cache managers update the base configuration; DeepSeek V4 supplies its custom layers and typical step while preserving the generic constraints.

Test Coverage

• Added and updated unit coverage for distributed resume headroom, generic estimation constraints, DeepSeek V4 typical-step configuration, and max_util_for_resume validation.
• pre-commit passed for all changed files.
• python3 -m py_compile passed for the changed Python files.
• Focused pytest execution was not available locally because the environment lacks compatible compiled tensorrt_llm.bindings.

PR Checklist

• 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, comment /bot help.