Customise nps.from_dtype() with keyword argument passthrough

[Zac-HD]

When this can also be hooked up to arrays(..., elements=kwargs), it's almost surprising that it took us so long to think of - it's reasonably easy to implement, conceptually obvious, and very useful to generate arrays of all-finite or bounded elements without falling afoul of our precision-loss and overflow checks.

@rsokl: note that I moved our existing tests for from_dtype() to a new test module, as test_gen_data was getting way too large. You should therefore probably review the two commits separately.

The extension for complex_numbers() would be almost trivial, except that the strategy doesn't currently accept a width argument and I would prefer not to add that into the same change as everything else here. Happy to add that on request from someone who would actually use it though!

Closes #2552.

[rsokl]

This is looking good! As commented below, we should have tests for the arrays(dtype, elements=mapping) use case, since this will likely become one of the most popular patterns among the extra.numpy strategies.

[rsokl]

Looks good!