[WIP] Tensor scale nvfp4

[nastya236]

Add tensor fp32 scale for nvfp4 quantization.
Python API changes:

• quantize returns tensor_scale if mode==nvfp4
• dequantize inputs tensor_scale if mode==nvfp4
• qqmm inputs 4 arrays if weights provided quantized and mode==nvfp4: quantized weights, scales and tensor_scale

Changes to the quantization:

• Quantize::eval_gpu quantize : 2 kernels launched: 1) all_reduce with AbsMax reduction 2) use computed absmax in the fp_quantize kernel
• Quantize::eval_gpu dequantize: inputs absmax and uses it in fp_dequantize.

Changes to qqmm:

• quantize as described above
• use tensor_amax_x and tensor_amax_w to compute alpha for BLAS operation. Also allocate beta = 0 because cublaslt requires both pointers to be on device or host.

It still looks ugly.

[awni]

This is a bit of a nit, but I don't really love the name tensor_scale. Partly because we call them array in MLX. What do you think about global_scale, or maybe array_scale?