perf(gdn): remove gdn prefill unnecessary h_state copy

[minedec]

Summary

Remove the previous redundant h-state conversion when persisting intermediate SSM states into the tracking pool (h-track and final-track) on the flashinfer GDN prefill fast-path. Instead of rebuilding an FLA-style h tensor from flashinfer's checkpoints, the tracking now indexes the native checkpoint buffer directly.

Approach:

• Reuse flashinfer's native checkpoints; drop the FLA-h rebuild: The prefill wrapper no longer runs a host-side per-sequence loop that splices the initial state into slot 0 and drops the trailing checkpoint to fabricate an FLA-style h tensor. It now returns the raw checkpoint buffer plus cumulative start offsets (checkpoint_cu_starts), and the caller indexes with flashinfer-native offsets. This removes a full-state allocation/copy and a host loop that incurred .item() syncs.
• Adjust index math to the new convention: Track indices now address the raw checkpoint buffer directly. The per-request entry count changes from ceil(L/C) to L//C, and the source index is shifted by -1 accordingly (FLA h[m] = ckpts[m-1]).
• Use a fused kernel for final-track copies: The final-state movement of conv_states/ssm_states switches from per-tensor index-copy to a fused Triton kernel, which is also extended to support 2D non-contiguous per-layer views, relax the contiguity requirement, and cast indices to int32 automatically.

Test Plan