feat(thor): add FA4 and stride-based split-N NVFP4 support for Wan2.2

[Michael20070814]

Summary

This PR adds the Wan2.2 acceleration path for NVIDIA Thor:

• enables NVFP4 kernels on Thor (sm_110a);
• adds a stride-based, batched split-N NVFP4 GEMM for both Wan FFN projections;
• integrates FlashAttention-4 variable-length attention and FA4-backed dynamic sparse attention;
• adds Thor configurations for NVFP4, split-N, and split-N with torch.compile.

Performance

For Wan2.2 MoE I2V on Thor (4 denoising steps, 81 frames, 480x832), the complete optimized configuration reduced end-to-end inference latency from approximately 45 s to 16 s, corresponding to a 2.81x speedup and a 64.4% latency reduction.

This result reflects the complete configuration and should not be interpreted as the isolated gain from split-N, NVFP4, or FA4.

Before the optimization:

After the optimization: