Fix sparse decode paged KV handling

[jasonlizhengjian]

Fix paged sparse-decode KV addressing needed by vLLM MiniMax M3
speculative verification.

• Propagate the actual K and V token strides to CUTLASS instead of assuming
  head_dim.
• Bound sparse-decode trip counts by the valid page count so short contexts do
  not load padded -1 page indices.
• Add regression coverage for interleaved FP8 paged KV and short contexts.

Validated on B300 through
vLLM #50032: 17 focused
TP1/TP4 parity, dispatch-gating, ragged-length, plan-reuse, and CUDA graph
replay cases pass. A TP4 end-to-end run also dispatched the kernel with 32
active speculative requests at a measured acceptance length of 2.81.