Add bf16 from from numpy.

[AWoloszyn]

Not submitting until I can actually find the documentation for 'g'.

[benvanik]

There's a few things you'll have to do, probably.

https://github.com/GreenWaves-Technologies/bfloat16/blob/97e0bab610569f5be550aee1526e94131905ac01/bfloat16.cc#L451 is a definition - there's no official support so it's a hack everywhere.

But matching with what pytorch does is the important part. You'll want to get a file dumped from there (as in pytorch/pytorch#143042) and add a test for it. iree/tooling/testdata/npy/ has the script that generates the npy files and numpy_io_test.cc has the tests that verify them (we may have to just check that in, as the script that generates in there is just plain numpy).