Object-based checkpointing support for unidirectional cuDNN LSTM cells

Object-based checkpointing support for unidirectional cuDNN LSTM cells

Once checked in, this will be the only way I know of to save canonical weights
when executing eagerly. Eager's name-based saving support will only do the
opaque parameter buffer.

I'm not going to try converting everything in one go, but it's a start at
least. And everything else should raise a NotImplementedError rather than
silently not saving correctly.

Single-layer cuDNN cells can be swapped for un-wrapped cuDNN compatible cells or
single cells wrapped in MultiRNNCells. Multi-layer cells need MultiRNNCell
wrapping.

PiperOrigin-RevId: 191905703