Implement metal kernel for GPUs on Mac

[casper-hansen]

MIT released their TinyChatEngine which includes kernels for every kind of platform. To ensure wide availability, we should integrate the kernels for Metal which are kernels for running quantized on MacOS.

Requirements:

• Run the Metal kernel through Python (maybe run on the fly?)
• Automatically switch to the Metal kernel if torch.backends.mps.is_available(), otherwise use CUDA.
• Figure out: How to support either GEMV/GEMM format with Metal

Kernel:
https://github.com/mit-han-lab/TinyChatEngine/blob/main/kernels/metal/kernel/op.metal#L10