Support fp8e5m2 dtype

[lsy323]

Add support for torch.fp8e5m2 dtype.

Right now torch_xla/csrc/tensor_util.cpp contains many duplicated codes, in the next PR I'll try to refactor it to make it cleaner.

The other fp8 variants and will be added in the following PRs

Test:
Added fp8 unit test

[miladm]

• Please include a documentation for FP8 - with example for users. Feel free to enrich the document as you add more FP8 capabilities/variants.

• How can we add a small micro-benchmark to the documentation to illustrate the benefits of FP8 vs. the baseline BF16?

• To my knowledge, PT AMP does not support FP8; correct?

[miladm]

@lsy323 we should include an accuracy correctness analysis in our benchmarking and documentation.

[JackCaoG]

approve to unblock, mostly lgtm we can have a follow up pr to fix review comments

[lsy323]

Hi @JackCaoG, @miladm, thank you for reviewing! Sorry for the delayed respond for the comments.

we should include an accuracy correctness analysis in our benchmarking and documentation.

In this PR the unit test is included to verify the correctness of the fp8 dtype and matmul op. Let's have documentation in a separate PR (After e4m3 variants are ready)