Tensor scale nvfp4

[nastya236]

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

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

Also some files related to qqmm were refactored.

Important details:

• [qqmm] currently if global_scale is provided for the first input, it must be provided for the second input as well. This is because we pass global scales as inputs in QQMatmul::eval_gpu() and we can't distinguish between global_scale_x and global_scale_w.
• alpha and beta both should be device or host ptrs. Therefore, if alpha is a device ptr,
  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)))

TODO:

• we probably want to support global_scale in metal as well but it requires changing all quantized operations
• it is not yet clear what might be the best strategy for scale computation during training (for x, w as well as for cotan), therefore QQLinear does not have global scale support. I will add it after some exploration as a separate PR.
• fp_qmv_impl was not updated yet to support global scale