Support float64 array about numpyro HOT 2 CLOSED

When the functionalities are ready, we should test for float64 array too.

Note that float64 will be significantly slower than float32 for GPU, so I think it is a good practice to write our code assuming that we will run it on float32, i.e. ensure numerical stability whenever possible. e.g. by providing logit support in distributions. I think tests passing with float32 should almost always pass with float64 (the reverse isn't true)?

from numpyro.

@neerajprad This is to prevent something like a = np.array([1.]), np.arange(...), or random.normal(...) without dtype in the code.

from numpyro.