Add support for weights in Empirical distributions

Hello,

If I'm not mistaken Empirical distributions don't allow for non-uniform weighting at the moment. This could be of tremendous value for any application using importance sampling. Is there any plan for supporting it in the future?

Ideally the signature should be left almost unchanged:

tfp.distributions.Empirical(
    samples, weights=None, event_ndims=0, validate_args=False, allow_nan_stats=True,
    name='Empirical'
)

Then the implementation of the different distribution methods would be similar to the ones you already have, for example instead of drawing indices uniformly in _sample_n you could simply draw them from a multinomial distribution or using standard resampling techniques such as systematic or stratified resampling to reduce the variance.

It seems to me like it would also help to solve this issue

Start at tfp.distributions.Empirical and review its existing distribution methods, especially _sample_n, along with the related issue #704. The work is done when Empirical accepts an optional weights argument and its distribution behavior uses those weights consistently.