Low-precision integer type support

Bandicoot does not currently support 8-bit or 16-bit integer types in matrices, which was an oversight that was never fixed previously because it would have caused too many kernels to be compiled.

But now that kernels are compiled on-demand, this isn't an issue anymore. So, this MR enables those types. Some small changes were necessary to the CUDA and OpenCL backends to make tests pass.

The huge number of changes in the tests directory is because, like ensmallen, I adapted the templated tests to use a macro that defines the types of interest for testing. I then adapted all of Bandicoot's tests accordingly (51k lines of tests... a lot of tedium).