Add FlashVSR chunked decoding for long videos avoid OOM

[anastasiuspernat]

Add FlashVSR context window support for long videos

Problem

FlashVSR video upscaling with context windows on long videos (500+ frames) causes:

1. OOM errors during decoding (115+ latent frames)
2. Visible transitions at chunk boundaries due to latent blending if directly using context window

Solution

1. New flashvsr fuse method (context_windows/context.py)
   No-blend mode for FlashVSR: overlap provides temporal context, but predictions are discarded (not blended).

• First chunk: weight=1.0 for all frames
• Later chunks: weight=0 for overlap, weight=1.0 for new frames

2. WanVideoDecode uses context_window settings passed from Wan Context Window options and does
   chunked decoding with overlap trimming (nodes.py)

• Decode chunks with overlap for temporal context
• Discard overlap frames (no blending)
• Key fix: Proportional overlap calculation accounts for FlashVSR's frames_to_trim=3 behavior (outputs 41 frames from 11 latent, not 44 if context_frames is set to 44)
  Calculation: actual_overlap = decoder_output × (latent_overlap / latent_context_frames)

4. Updated WanVideoContextOptions
   Added flashvsr to fuse_method dropdown

Usage Example specific for FlashVSR

  context_schedule: "static_standard"
  context_frames: 44  # Adjust for VRAM
  context_overlap: 16
  fuse_method: "flashvsr"  # NEW