feat (quant/mx): Added midmax scale rounding option to MX types

[nickfraser]

Add "midmax" scaling for MX datatypes. Midmax is a rounding mode for the shared scale in MX datatypes. Also, plugs MidMax scaling into the LLM example. The standard mode (as referenced in the OCP MX datatype spec) is "floor" and does the following:

$$ po2\_shared\_scale = \lfloor log_2 (\| x \|_\infty) \rfloor $$

Midmax replaces this floor operation with a special rounding mode that reduces the rounding error in the maximum value in $x$, the the cost of potentially increasing the rounding error in the smallest values in $x$.

Rerunning the experiments from the Post-Training Model Expansion paper, we get the following results:

model	spinquant	expansion_step	scale_round_func	float_ppl	quant_ppl	ARC-C	ARC-E	HS	WG	PIQA	all_acc
meta-llama/Llama-3.2-1B	False	0	floor	8.938	11.694	0.289	0.597	0.418	0.559	0.693	0.511
meta-llama/Llama-3.2-1B	False	0	midmax	8.938	11.574	0.289	0.610	0.425	0.556	0.712	0.519
meta-llama/Llama-3.2-1B	False	7	floor	8.938	11.452	0.283	0.611	0.426	0.575	0.701	0.519
meta-llama/Llama-3.2-1B	False	7	midmax	8.938	11.241	0.272	0.619	0.430	0.574	0.701	0.519
meta-llama/Llama-3.2-1B	True	0	floor	8.938	11.518	0.305	0.628	0.422	0.569	0.709	0.527
meta-llama/Llama-3.2-1B	True	0	midmax	8.938	11.552	0.293	0.595	0.424	0.562	0.707	0.516
meta-llama/Llama-3.2-1B	True	7	floor	8.938	11.359	0.298	0.590	0.433	0.556	0.697	0.515
meta-llama/Llama-3.2-1B	True	7	midmax	8.938	11.294	0.303	0.606	0.430	0.559	0.699	0.519

Note that beyond the OCP MX v1 spec datatypes, MidMax is not thoroughly tested and should tested further when looking beyond these types.

Also, while adding this feature, I took the opportunity to remove some duplicated code between MXWeightMixin and MXActMixin into a parent class (MXMixin).

[nickfraser]

1 comment, otherwise ready for review.

[Giuseppe5]

One small change, then it can be merged.