Describe the 'float32_matmul_precision' settings in more detail (#107169)
The documentation for `torch.set_float32_matmul_precision()` mentions a datatype called "bfloat16_3x". This doesn't appear to be a very standard term, and I had a hard time figuring out what exactly it meant. I now assume it refers to [[Henry2019]](http://arxiv.org/abs/1904.06376), which describes an algorithm by which a float32 multiplication is approximated via three bfloat16 multiplications. This PR updates the documentation to include this reference and to briefly describe how this algorithm works.
Note that I just learned everything that I wrote here, so I'd appreciate if someone more expert in this topic could check to make sure that I didn't get anything significantly wrong.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107169
Approved by: https://github.com/colesbury
