Add a `mask` argument to `RNN.__call__`

Some modules that receive a `mask` argument in their `__call__` method include:

* [BatchNorm](https://flax.readthedocs.io/en/latest/api_reference/flax.linen/layers.html#flax.linen.BatchNorm)
* [SelfAttention](https://flax.readthedocs.io/en/latest/api_reference/flax.linen/layers.html#flax.linen.SelfAttention.__call__)
* [MultiHeadDotProductAttention](https://flax.readthedocs.io/en/latest/api_reference/flax.linen/layers.html#flax.linen.MultiHeadDotProductAttention)

This is useful because it allows one to dynamically exclude elements from a reduction.

**Feature request:** Add a `mask` argument to [RNN](https://flax.readthedocs.io/en/latest/api_reference/flax.linen/_autosummary/flax.linen.RNN.html#flax.linen.RNN.__call__)'s `__call__` method that makes the RNN's carry skip over masked-out elements. Thus the RNN behaves as if the masked-out elements aren't there. This can be implemented by using a _masked scan_ function. An example implementation of the latter can be found at https://github.com/google/jax/issues/18664.