MX-scale discrepancy during quantization and dequantization

[mariosfourn]

In the case of very small numbers input numbers around the subnormal range of torch.float or torch.bfloat16, the scale exponent will take its smallest unbiased value: -127. However, you only allow division with a scale of 2**-126 in line 143 of mx_tensor.py. This is because of an incompatibility with triton.

However, during dequantization you use the the smaller scale of 2**-127 when calling

s_fp = get_fp_scale(scale_e8m0).reshape(-1, 1).to(target_dtype

in line 235. Why not clip the exponent to -126 in the get_fp_scale function

[jainapurva]

@drisspg @vkuzo