perf(pd): compact token forwarding packets

[WANDY666]

Summary

• Encode the common no-logprobs PD token update as a fixed tuple containing only fields consumed by PD master.
• Preserve current main metadata, including MTP accepted/verify counters, first-token node mode, and multimodal input usage.
• Fall back to the legacy packet whenever optional metadata is present.
• Split mixed compact/legacy batches into contiguous wire packets while preserving token order.
• Reconstruct the existing metadata and FinishStatus contract at PD master.

Dependency

This is a stacked PR on perf/pd-omit-unused-logprob-metadata (#1486), because that PR supplies the request-level marker selecting the common no-logprobs path. After #1486 merges, this PR can be retargeted to main without carrying unrelated changes.

Benchmark context

The historical cfce087 32K microbenchmark measured packet size around 8.4 KB → 3.6 KB, D-side serialization around 168 µs → 19 µs, and PD-master deserialization around 868–1366 µs/packet → 32 µs. The full q15 result remained within run-to-run noise, so this is presented as CPU/transport headroom rather than a proven QPS-capacity increase.

Validation

• Production-code diff passes repository black and flake8 hooks.