fix(pd): account for KV transfer pages in mem_fraction

[sufubao]

Summary

• reserve the future paged KV transfer buffer when auto-profiling token KV capacity on Prefill and Decode workers
• share one buffer-shape calculation between memory profiling and the actual transfer allocation, including the DeepSeek MLA layout
• fail early with actionable configuration guidance when the reservation leaves no token KV capacity

Motivation

The PD transfer process allocates its GPU page buffer after model initialization and KV-cache profiling, so mem_fraction could consume that space before the transfer process starts. With Qwen3.5-27B, 16 pages, page size 4096, 17 full-attention KV layers, 4 global KV heads, head dimension 256, and BF16, the unprofiled allocation is 4.25 GiB per GPU.

This change subtracts the exact allocation size from the automatic KV-cache budget. Normal and PD-master modes are unchanged, and an explicit max_total_token_num continues to bypass automatic profiling.

Validation

• Black 21.12b0 check passed on modified files
• flake8 6.1.0 passed on modified files
• compileall and git diff --check passed
• no new unit tests are included

Validation note

ACP GPU end-to-end validation was not run from the local macOS development environment.