Fix DSv4 MTP graph metadata diagnostics

[bhaktatejas922]

Summary:

• fix DSv4 EAGLE/MTP decode metadata handling during CUDA graph capture/replay
• add opt-in DSv4 graph-capture tracing around collectives and CUDA graph runner phases
• add opt-in B12X sparse MLA shadow-compare diagnostics for FlashMLA reference checks
• skip DSv4 sparse MLA shadow compare during CUDA graph capture using both torch stream-capture state and SGLang capture context
• enrich shadow records with capture state, forward context, valid/padding row split, and row-wise compare stats so post-capture traffic can be separated from graph dummy rows
• keep B12X sparse MLA draft-decode gated off by default while target-verify and draft-extend remain explicit opt-in paths

Latest validation:

• python3 -m py_compile python/sglang/srt/distributed/parallel_state.py python/sglang/srt/layers/attention/deepseek_v4_backend.py python/sglang/srt/model_executor/cuda_graph_runner.py
• Full TP4 C32 all-mode sparse gate started with CUDA graphs enabled and B12X sparse MLA hits increasing, but throughput collapsed; saved logs showed EAGLE draft graph replay padding around raw_bs 29-31 to graph_bs 32

Notes:

• The previous shadow mismatch records from dsv4-c32-shadow-20260514T130244Z were capture dummy rows, not real traffic. This PR now prevents that class of false diagnostic.
• Next investigation should focus on EAGLE draft/target-verify metadata and padded replay rows before promoting all-mode sparse.

[bhaktatejas922]

@claude review this pr