Support more batch distributions in HaarReparam

[fritzo]

This adds support for more uses of HaarTransform, motivated by batched epidemiological models.

1. Documents the experimental_enable_batch kwarg (as suggested by @martinjankowiak)
2. Mentions the experimental_enable_batch kwarg in the relevant error message.
3. Adds reshape_batch() support for TransformedDistribution, ComposeTransform, and _InverseTransform, which are needed by HaarReparam in my model.
4. Relaxes checking in block_plate to not error if no matching plate is found.
5. Adds support for dist.Uniform
6. Works around a new error in torch.view_as_complex()

Tested

• added unit tests for reshape_batch()
• tested end-to-end in an epidemiological model

[fritzo]

Thanks for reviewing!