LSTM SymInt-aware changes & meta registration (non-cuDNN CUDA) (#90701)
Adds meta registrations for cuDNN and vanilla CUDA ops underneath `lstm()` and makes the logic SymInt-aware.
TODO:
* cuDNN side does some [nasty stuff](https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/cudnn/RNN.cpp#L1567) with buffers; this needs larger redesign to figure out
* Indicate that AOT Autograd can be used when an LSTM is present (remove the check for this once it's fully supported)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/90701
Approved by: https://github.com/ezyang
