Tensor scale nvfp4

[nastya236]

1. Add per tensor scale for nvfp4 quantization for cuda and cpu.

qqmm, quantize, dequantize inputs optional 1D float32 array (global_scale) if mode == "nvfp4".

Tensor scale will help with small inputs:

import mlx.core as mx

x = mx.random.uniform(shape=(2, 16)) / 1e5
xq_ns, scales_ns = mx.quantize(x, mode="nvfp4")
global_scale=mx.absmax(x).astype(mx.float32)
xq_s, scales_s = mx.quantize(x, mode="nvfp4", global_scale = global_scale)

print(mx.allclose(scales_ns, mx.zeros_like(scales_ns)))
print(mx.allclose(scales_s, mx.zeros_like(scales_s)))

2. AbsMax reduction type and mx.absmax op

For nvfp4 training we will compute amax often. So now there is a new reduction type in which abs is applied inside the all_reduce kernel.
Probably there is a way to do it better.

x.shape() = (4*4096, 11008)
mx.absmax(x): 0.000166 s
x.abs().max(): 0.000284 s

TODO: we probably want to support global_scale in metal as well but it requires changing all quantized operations.