Fix FlashVSR attention mask and output quality

- Use generate_draft_block_mask_refined for sparse attention mask (matches
  naxci1's generate_draft_block_mask_sage with proper half-block key scoring)
- Remove spurious repeat_interleave(2, dim=-1) from generate_draft_block_mask
  that doubled the key dimension incorrectly
- Add torch.clamp(0, 1) to _to_frames output (matches naxci1's tensor2video)
- Add .to(self.device) on LQ video slices in streaming loop for all pipelines

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

inference.py

@@ -805,7 +805,7 @@ class FlashVSRModel:
         from einops import rearrange
         v = video.squeeze(0) if video.dim() == 5 else video
         v = rearrange(v, "C F H W -> F H W C")
-        return (v.float() + 1.0) / 2.0
+        return torch.clamp((v.float() + 1.0) / 2.0, 0.0, 1.0)
 
     # ------------------------------------------------------------------
     # Main upscale method