[perf]: hand the VSA gate to video_sparse_attn as a transposed view

The 64-tile BHSD path copied gate_compress into a contiguous BHSD tensor on
every call. The coarse/sparse combine views the gate at block resolution,
which never needs contiguity, so the transposed BSHD view is passed as is.
This saves one full-sequence copy per attention layer with bit-identical
output. On a GB200 at 1x39936x12x128 (Wan 1.3B, 480p) the per-call peak
drops from 639.8 to 521.8 MiB and latency from 11.64 to 11.29 ms.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NjFimeedTQWxgqSzP5xo4j

Author: SolitaryThinker

fastvideo/attention/backends/video_sparse_attn.py

@@ -327,11 +327,14 @@ def forward(  # type: ignore[override]
 
         if video_sparse_attn is None:
             raise NotImplementedError("video_sparse_attn is not installed")
-        # Default 64-element-tile path (unchanged): BHSD round-trip.
+        # Default 64-element-tile path (unchanged): BHSD round-trip. The gate is
+        # only read elementwise by the coarse/sparse combine, which views it at
+        # block resolution without requiring contiguity, so the transposed view
+        # is handed over as is (no full-sequence copy).
         query = query.transpose(1, 2).contiguous()
         key = key.transpose(1, 2).contiguous()
         value = value.transpose(1, 2).contiguous()
-        gate_compress = gate_compress.transpose(1, 2).contiguous()
+        gate_compress = gate_compress.transpose(1, 2)
         return video_sparse_attn(query,
                                  key,
                                  value,