[None][fix] Correct MiniMax-M3 Eagle draft sub-page pool bound

[zheyuf]

What this fixes

This PR fix the cold start AL regression for Minimax-m3 eagle. It's a functionality fix but not a perf boost. No extra overhead is added.

MiniMax-M3 uses a shared KV cache for the target model and the Eagle3 draft model. The target uses 128-token pages, while the draft model views the same memory as 32-token pages.

During cold start, the draft view reported the wrong number of available cache pages. Some valid draft pages could therefore be treated as out of range, causing the draft model to read incorrect KV data. The target model still verified the output, so final answers remained correct, but Eagle acceptance length dropped to roughly 1.8-2.0 and speculative decoding became less effective.

What changed

The draft view now calculates its cache boundary from its own 32-token layout instead of reusing the target manager's 128-token value.

This PR only fixes that boundary calculation. The aggregated shared-KV routing and performance changes are handled separately in #17835.

Result

Three independent cold starts recovered first-round acceptance length:

cold start	acceptance length
1	3.264
2	3.417
3	3.489

Mean acceptance length was 3.390. Cold generation time improved from 13.407 s to 10.946 s on average, an 18.4% reduction. Hot runs showed no regression.

Validation

• Three independent two-node cold starts completed successfully.
• 288/288 long-prompt requests completed.
• No traceback, assertion, CUDA, segmentation-fault, or fatal errors were found.
• All workers used the expected fixed source.
• Focused tests: 13 passed.
• Formatting, static checks, and DCO passed.

PR checklist

• PR title follows the [JIRA/NVBUG/None][type] format
• Commit is signed off (DCO)
• Focused unit tests pass
• Cold-start regression reproduced before the fix and passed after it

[zheyuf]

/bot run

[tensorrt-cicd]

PR_Github #66836 [ run ] triggered by Bot. Commit: b09f72c Link to invocation

[brb-nv]

LGTM.

[tensorrt-cicd]

PR_Github #66836 [ run ] completed with state SUCCESS. Commit: b09f72c
/LLM/main/L0_MergeRequest_PR pipeline #54396 completed with status: 'SUCCESS'

CI Report

Link to invocation