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

[mesaleh]

Replaces #260 after it was automatically closed for inactivity. GitHub is not allowing the closed PR to be reopened through the API, so this keeps the same fix alive on a fresh rebased branch.

Summary

• Keeps spec-decode CUDA graph capture seq_lens, output cache slots, and dummy KV page internally consistent during warmup and capture.
• Uses the reserved padding page for TRT-LLM MLA CUDA graph capture block tables.
• Flattens multi-token TRT-LLM MLA decode for both target verification and draft catch-up so each token gets its own causal seq_len bound.

Validation

• python3 -m py_compile python/tokenspeed/runtime/execution/cuda_graph_wrapper.py python/tokenspeed/runtime/layers/attention/backends/trtllm_mla.py
• git diff --check