Support f16, f128

[ExpHP]

numpy has float16 and float128 datatypes, which can appear in npy files.

There exists a crate called half for binary16-encoded floats.

• Support f16 (thanks @KeKsBoTer)
• Support f128

[ExpHP]

Some more oddities: numpy/numpy#10288

Basically, float128 is not binary128 format. It is for 80-bit precision floats.

• Supposedly there exists np.float96. I don't have this on my system.
• The dtypes np.float96 and np.float128 both represent np.longdouble, just with different amounts of zero-padding. (float96 aligns to 32-bit boundaries, float128 aligns to 64-bit boundaries).

Also, I had no idea this was the case, but: Even my own system has 80-bit precision for long double!

> np.finfo(np.longdouble).eps
1.084202172485504434e-19

[ExpHP]

Reopening because we still lack f128 support