[RFC] Add Float8E4M3FNUZ and Float8E5M2FNUZ to StableHLO.

[jakeh-gc]

This is a proposal to add Float8E4M3FNUZ and Float8E5M2FNUZ floating point types to StableHLO.
Feedback welcome, see rfcs/20230321-fp8_fnuz.md for more details.

[burmako]

@jakeh-gc Thank you for your RFC! Given its similarity to the F8E4M3FN and F8E5M2 RFC as well as the F8E4M3B11FNUZ RFC, I think that we have a strong case for the dtypes proposed in this RFC as well.

Similarly to the latter RFC which we have just approved, let's keep this one open for two weeks, i.e. until Wednesday 4/5, and then approve as well if no blocking feedback from the community.

[sherhut]

As stated in #1308 (comment), adding fp8 formats supported by hardware makes sense to me. The question about longer term remains. @burmako how would we organize work to explore address this? @jakeh-gc have you thought about precision_config?

Just to be clear: I am not asking for this longer-term design questions to block addition of the type itself.

[cantonios]

Apart from minor comment/question, LGTM.

[jakeh-gc]

@sherhut

@jakeh-gc have you thought about precision_config?

I'm not sure how precision_config should be dealt with. I'm not personally a fan of how it's defined at the moment:

* `DEFAULT`: Fastest calculation, but least accurate approximation to the
  original number.
* `HIGH`: Slower calculation, but more accurate approximation to the
  original number.
* `HIGHEST`: Slowest calculation, but most accurate approximation to the
  original number.

For an implementation, DEFAULT and HIGH seem like a license to do anything. I hope #755 resolves this 🤞.

My approach has been to at least use the precision of the input types, but that's certainly not endorsed by all of my colleagues who want "benchmarks" to go fast.

[burmako]

Thank you, everyone, for the lively discussion! It's been two weeks, and we have broad support for adding the 2 new dtypes, with all the feedback addressed. The RFC is now approved.

A recurring theme in the discussions of the last few weeks has been working out a long-term story for the fp8 dtypes. I think this is important to follow up on, and I'll work with @theadactyl on including this on the agenda of our community meetings.