[AnyPrecision optimizer] - needs unit tests

[lessw2020]

This is just an issue to track that we do not have unit tests added yet for the BF16 optimizer.

Main tests would be:
a - verify AdamW and BF16 set to FP32 dtypes produce matching results. (BF16 degrades to AdamW w/o Kahan and dtypes to FP32).
b - simple checks for confirming internal dtypes match expected (i.e. set variance to BF16 but momentum to FP32, verify both are in proper dtype).