Option to avoid bijector caching

[alexhepburn]

The current behaviour for tensorflow_probability.bijectors.Bijector is to cache the input to be used if the inverse function is called. I would very much like a keyword being able to turn this caching off to ensure the inverse function of the bijector is always called. Whilst setting the property of _is_injective can accomplish this, you lose other features.

It would be nice to have a parameter for bijector which could avoid using caching and instead explicitly use the inverse function at all times.

Thanks!