Allow coercion of Julia types to NumPy dtypes

[hameerabbasi]

Is your feature request related to a problem? Please describe.
We're using Julia types as dtypes, but a third-party library does np.dtype(jl_type). Maybe there's a way to support this as an upstream feature for everyone.

Describe the solution you'd like
Ideally, I'd like something like an implementation of the __numpy_dtype__ protocol or supported Julia types. This could even be implemented on some kinds of structs.

Describe alternatives you've considered
Maybe we could solve this upstream by converting back and forth, but a simple conversion is always helpful.

Additional context
Originally reported by @ilan-gold and @amalia-k510 in pydata/sparse#915

[cjdoris]

TIL about __numpy_dtype__, it looks like a recent addition. We could certainly implement it for Julia types (jl.Int64.__numpy_dtype__ being np.int64 etc).

[hameerabbasi]

That'd be great! I'm not super well-versed in Julia; but if the relevant code is in Python, I'm happy to make a PR.

Also small correction: jl.Int64.__numpy_dtype__() would be np.dtype(np.int64)

This has two minor differences:

1. The return must be a dtype instance, not a type like np.int64.
2. The protocol requires a callable, not an attribute directly.

[cjdoris]

It's OK I've done it (#738) - it was surprisingly simple because I could re-use existing type-conversion machinery from the array interface code. This means we can support struct types for free. It also means that doing np.dtype(eltype(julia_array)) should be consistent with np.array(julia_array).dtype in all cases.