Add M3 Mac compatibility with Metal Performance Shaders support

[zettai-seigi]

No description provided.

[SplittyDev]

Does this actually work for you?

1. The get_cuda_free_memory_gb function is still imported in demo_gradio.py, causing it to immediately crash on launch due to CUDA not being available. This is an easy fix: Simply remove the import, since it's unused anyway.
2. The dtype in the sample_hunyuan call is wrong (bfloat16, but expected half) and needs to be changed from torch.bfloat16 to transformer.dtype in order for the program not to crash during inference.

After making these adjustments, the model at least runs and inference does do something, but the output is completely black and performance is horrible (37.7s/it for me).

That said, I've been testing this on an M2 Ultra with 128GB unified memory. Not sure if it's different on M3.

[cachatj]

I was going to try this, vbut I am already up and running on MBP M3max - are we expecting performance improvements here or something?

while it's running, was this PR designed to address the error NotImplementedError: The operator 'aten::avg_pool3d.out' is not currently implemented for the MPS device.

[loscrossos]

i ported it to mac

https://github.com/loscrossos/core_framepackstudio