Add Turing safe-FP16 execution path

[Ownistic]

Summary

Adds a native safe-FP16 execution path for MiniMax H3 on pre-Ampere NVIDIA GPUs, including Turing.

The implementation is adapted for WanGP from the reference script:
https://github.com/Amduraznak/minimax-h3-fp16-fix/blob/main/minimax_h3_fp16_fix.py

Changes

• Enable the safe-FP16 path only for CUDA GPUs below compute capability 8.0 when H3 runs in FP16.
• Keep numerically sensitive residual and pointwise operations in FP32.
• Run compatible projections and attention operations in FP16 to reduce Turing denoising time without creating a duplicate full-FP32 model.
• Keep Qwen3-VL prompt encoding on SDPA while allowing the H3 transformer to use its selected attention backend.

Benchmark

Measured on an RTX 2060 Laptop GPU with 6 GB VRAM (SM 7.5) for a 5s video at 320p resolution:

• Before: approximately 150 seconds per H3 denoising step
• After: 31.9 seconds per step with the safe-FP16 path
• Result: approximately 4.7x faster, or 79% less denoising time