Implement cast support for ubyte and half

[juntyr]

Fixes numpy/numpy-quaddtype#27

[juntyr]

@SwayamInSync Does the version of C++ we use already have float16 support / what does numpy use so that I could also add casting support for float16?

[ngoldbaum]

Here's where it happens in the stringdtype prototype:

numpy-user-dtypes/stringdtype/stringdtype/src/casts.c

Line 1123 in b4a9429

STRING_TO_FLOAT_RESOLVE_DESCRIPTORS(float16, HALF)

You need to link against libnpymath to get the necessary C API:

https://numpy.org/doc/stable/reference/c-api/coremath.html#half-precision-functions

Here's where I handled that in the meson config:

numpy-user-dtypes/stringdtype/meson.build

Line 19 in b4a9429

npymath_path = incdir_numpy / '..' / 'lib'

[juntyr]

The other problem, and probably why ubyte wasn't added earlier, is that ubyte and bool alias but need different code paths. Not sure how to fix that ...

[juntyr]

@SwayamInSync @ngoldbaum I think everything should work now, what do you think?

[SwayamInSync]

LGTM

[ngoldbaum]

Thanks for doing this @juntyr!

[SwayamInSync]

BTW if needed then in this SwayamInSync#13 I wrote the workaround to get an "Non-Implemented" error instead of segfault crash.

[ngoldbaum]

Ah, I missed that - can you send it in as a followup PR?