It seems that np.dtype(...) does not produce consistent hashes and does not equal the hash of the contained dtype. I feel like these should be consistent with each other so that the following will return true:

np.empty(1).dtype in {np.float32, np.float64}  # False

Reproducing code example:

In [1]: np.float64 == np.dtype(np.float64)
Out[1]: True

In [2]: np.float64 in {np.dtype(np.float64)}
Out[2]: False

In [3]: hash(np.float64)
Out[3]: 8793996338852

In [4]: hash(np.dtype(np.float64))
Out[4]: 3855163183283070272

hash(np.dtype(np.float64)) returns a different value when called in different runtimes too.

Would this be worth fixing?