fix(trtllm-mla): make spec-decode CUDA graph capture causal

[mesaleh]

Summary

• initialize spec-decode CUDA graph capture with internally consistent synthetic sequence lengths and cache slots
• point TRT-LLM MLA capture block tables at the same synthetic per-request pages and zero those pages before capture
• flatten multi-token TRT-LLM MLA decode into per-token decode queries so target verification and draft catch-up use causal sequence bounds per token

Motivation

While bringing up Kimi K2.6 NVFP4 with EAGLE3 speculative decoding and --attention-backend trtllm_mla, CUDA graph startup could enter an impossible synthetic state: multi-token decode capture used grouped metadata that did not match the dummy KV/cache locations. This led to capture-time instability and downstream invalid draft ids.

This PR is intentionally independent of #217. It does not touch the EAGLE first-step reduce plumbing and merge-tests cleanly with pull/217/head.

Validation

• python3 -m compileall -q python/tokenspeed/runtime/execution/cuda_graph_wrapper.py python/tokenspeed/runtime/layers/attention/backends/trtllm_mla.py
• Local GB200 validation: Kimi K2.6 NVFP4 + EAGLE3 + trtllm_mla captured batch sizes [1, 2, 3, 4, 5, 6, 7, 8] and reached healthy readiness
• OpenAI-compatible synthetic benchmark completed 10/10 requests with p50 TPOT 3.760 ms and mean decode 264.676 tok/s