Implement ndarray.byteswap

[bmerry]

It's implemented for numbers (including complex) and boolean.

On a 1D array it's about 7x faster than numpy. For higher dimensions it is slower, and it is much slower for Fortran-order arrays, for the same reasons as in #9400.

This is almost my first PR for numba and I'm pretty new to LLVM as well, so I've got a few questions:

• I've kept the code smaller by unconditionally bit-casting to and from a suitably-sized integer type. If it's already an integer type (probably the common case) that's unnecessary LLVM IR instructions, but I'm assuming that's harmless.
• In numpy, byteswap on a scalar with inplace=True raises an exception, and I've mirrored that behaviour. Since numba is statically typed, an alternative is to simply not support that useless parameter, which would move the error to compile time (generally a good thing) but be less compatible with numpy. Any thoughts?
• Is there a better way to handle the array iteration? I looked into wrapping the @intrinsic scalar function into a ufunc and calling that (which uses a sensible iteration order for Fortran-order arrays), but ran into some cyclic import problems.

[bmerry]

Also, is there anything that should be added/tested for CUDA? I've never tried out the CUDA support in numba.

[stuartarchibald]

Thanks for submitting the PR @bmerry, it's been queued for review. RE: CUDA, there's nothing to be added for now, there is much less NumPy support on the CUDA target than on the CPU target, if the need arises it can go into a separate PR. Thanks again!