mps: add nf4 dequantize/quantize kernel

[bghira]

this ports the CUDA NF4 support to Metal.

so far, I've targeted nf4 quant/dequant because it's one of the least-accessible formats for Mac users.

we're using uint8 under the hood. for what it's worth, Metal (and the underlying hardware) lacks fp8/fp4 support.

performance has not been the forefront of this effort, as most of the time was spent determining how to plug metallib into bitsandbytes and correctly build it.

I'd like some feedback on this approach, because due to my inexperience with your build toolchain, it's highly likely I've done things in ways that can be improved.

I'm building on lessons I'd learnt while building a pytorch custom op for universal-metal-flash-attention, namely the way the MTLBuffers are retrieved from torch MPSGraph objects, which required the use of the torch headers.

[bghira]

@rickardp cc