RNN / GRU / LSTM implementation for torch_xla

[qihqi]

🚀 Feature

Given that the experimental launch of scan operator that lowers to XLA's WhileOp, we should leverage it to implement performant RNN layers.

It should great to leverage it in place of the for loop that loops throught the time dimension, which could be large.

Motivation

for loops defined in python will be traced through by Lazy tensor and create a huge XLA graph. The runtime of these graphs might even be better than the rolled loop; but the compile time would be too long.

Scan operator is the technique for shortening compile time, this is also used in flax's RNN implementation.

Pitch

Alternatives

Additional context

[qihqi]

Scan feature introduced here: https://cloud.google.com/blog/products/application-development/pytorch-xla-2-6-helps-improve-ai-model-performance

[qihqi]

related: #2422