feat: implement LSTM and GRU operators for torchlib (#2674)

Implement aten_lstm and aten_gru operators to enable torch.onnx.export
for PyTorch LSTM and GRU layers. This addresses issue #2546.

Key features:
- Full support for multi-layer RNNs (num_layers > 1)
- Bidirectional support (forward and backward directions)
- Handles both biased and non-biased configurations
- batch_first parameter support with automatic transposition
- Dropout support between layers (nondeterministic seeded)
- Proper gate reordering for ONNX compatibility:
  * LSTM: PyTorch [i,f,g,o] -> ONNX [i,o,f,g]
  * GRU: PyTorch [r,z,n] -> ONNX [z,r,n]

Implementation details:
- Uses ONNX LSTM/GRU operators with proper parameter formatting
- Handles weight matrix transposition and reshaping
- Correctly concatenates biases using op.Concat
- Processes each layer independently with proper state management
- Returns outputs in PyTorch-compatible format

Closes: #2546

Also resolves:
- pytorch/pytorch#120626 (GRU export)
- pytorch/pytorch#123089 (LSTM export)  
- pytorch/pytorch#164834 (GRU dynamo export)

Author: ombrdr47