Handling underflowing denorms in float to half conversion for `muamus_float`

[fangq]

As reported by Kris in this mailing list post
https://groups.google.com/g/mcx-users/c/WVv0XYpJx9M/m/sUZnIFdnCAAJ

mcx's float->half packing code fails in some input data range.

after debugging, the failed values fall between 2^-16 (1.5259e-05) and 2^-14 (6.1035e-05), which are known as denorms (or subnorms).

need to add specialized handling for input of this range.