Pass **kwargs through npst.from_dtype() to the inferred strategy

[Zac-HD]

Consider the following pattern:

arrays(floating_dtypes(), shape, elements=floats(-10, 10, width=...))

If width=None, when the generated dtype is float32 we will trigger the warning for inexact representation of element values (which is useful in general, if annoying here). Alternatively, if width=32 then we have massively reduced the precision of the element values that we can generate when the dtype is float64. There are of course workarounds with e.g. flatmap in current use, but they're pretty ugly and awkward for new users.

I therefore propose that:

• npst.from_dtype() should take **kwargs, and pass them through to whichever strategy function it chooses. In practice this would usually just be min_value and max_value for integers or floats, but using **kwargs gives a cleaner implementation and allows use with timestamps, complex numbers, string length, and so on.
• The arrays() strategy would then take elements: Union[SearchStrategy, Mapping[str, Any]] = None. The mapping case would then be used as kwargs for a call to from_dtype().

CC co-designer @rsokl.

[Zac-HD]

master...Zac-HD:dtype-passthrough is working! I'm going to hold off on a PR until after #2615 and #2613 are merged though, because I don't need any more merge conflicts at the moment 🙂