support asymmetric pd-mtp via mock spec hidden

[hnyls2002]

Always size the PD MetadataBuffers hidden_states buffer to model_config.spec_hidden_size on both prefill and decode (instead of conditional on local --speculative-algorithm). This unblocks asymmetric P/D configs where prefill runs no spec module but decode runs EAGLE/MTP: prefill ships a zero-initialized buffer that decode consumes as mock conditioning for the first draft step. Verified spec decoding keeps the output token correct -- only the first decode iteration's accept length is affected (~ < 1% throughput hit on long generations).

Cost: when neither side runs spec, the buffer is allocated but unused (~few MB). The alternative is a wire-protocol size mismatch on asymmetric configs.

Test: test_dsv4_pd_disagg_nixl.py uses asymmetric config (prefill TP-only, decode TP+DP+EAGLE), no new server arg required. Follow-up to #23918.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!