[None][fix] Split draft host cache during estimation

[zheyuf]

@coderabbitai summary

Description

When one-engine speculative decoding creates separate target and draft KV cache managers, host_cache_size is a combined budget. The normal manager-construction path splits that budget by target/draft bytes per token, but the split was nested under the not estimating_kv_cache guard.

The estimation path still creates both host pools, so each manager temporarily received the full configured host budget. With a 192 GiB per-rank budget at TP8, startup could request up to 3 TiB of host memory on a 2 TiB node before reaching the later, correct split.

This change:

• splits the host budget whenever separate draft and target managers are used, including estimation;
• keeps GPU-budget splitting disabled during estimation, preserving the existing max-token sizing behavior; and
• leaves the caller-owned base config unchanged.

The issue affects host offload with separate speculative draft/target KV managers; non-speculative models, shared-cache paths, and configurations without host cache are unaffected.

Test Coverage

• Added test_estimation_splits_host_budget_before_creating_managers to cover the previously missed build_managers(..., estimating_kv_cache=True) lifecycle.
• python -m pytest -q tests/unittest/_torch/executor/test_kv_cache_budget_split.py: 21 passed in the branch-m3-2026-08-12 nightly environment with this source file overlaid.
• Ruff, YAPF, legacy lint baseline, and git diff --check pass for the modified files.
• Bia TP8 Eagle3 PDL-off/on runtime smoke with the original 192 GiB per-rank setting: Slurm array 751970 (queued; results will be added to this draft).

PR Checklist

• PR description clearly explains what and why.
• PR follows the TRT-LLM coding guidelines to the best of my knowledge.
• A regression test is provided for the new code path.
• No API changes or new dependencies.
• No CODEOWNERS, documentation, or architecture-diagram update is needed.

GitHub Bot Help

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