feat: add LSTM, GRU, and RNN ONNX node converters

[nik875]

Closes #136

Summary

This PR adds ONNX-to-PyTorch converters for the LSTM, GRU, and RNN operators, covering opset versions 7 and 14.

All three operators are registered via @add_converter in a new onnx2torch/node_converters/rnn.py module and wired into
init.py. operators.md is updated to mark them as supported.

Implementation notes

LSTM wraps nn.LSTM. ONNX gate order [I, O, F, C] is reordered to PyTorch [I, F, G, O] at conversion time by slicing and
re-concatenating weight chunks. Initial hidden/cell states are supported as optional graph inputs. sequence_lens, peephole
weights, input_forget=1, clip, and layout=1 raise NotImplementedError.

GRU uses a manual per-timestep loop rather than wrapping nn.GRU. This is necessary because the ONNX default
(linear_before_reset=0) computes:

ht = tanh(Wh·x + Wbh + Rh·(r⊙h) + Rbh)

whereas PyTorch nn.GRU implements linear_before_reset=1:

ht = tanh(Wh·x + Wbh + r⊙(Rh·h + Rbh))

These are mathematically different for non-zero hidden states. The manual loop correctly implements both formulas; weights are
stored as buffers in ONNX gate order [Z, R, H] without reordering. sequence_lens, clip, and layout=1 raise
NotImplementedError.

RNN wraps nn.RNN. No gate reordering is required. The activations attribute is respected for Tanh (default) and Relu.
sequence_lens, clip, and layout=1 raise NotImplementedError.

All three converters require W and R to be graph initializers (static weights). Bidirectional and reverse directions are
supported for all three operators.

Output shape

PyTorch recurrent layers return Y with shape [seq, batch, num_directions * hidden_size]. The converters reshape and permute
this to the ONNX-expected [seq, num_directions, batch, hidden_size].

Tests

32 new test cases in tests/node_converters/rnn_test.py using the standard make_model_from_nodes + check_onnx_model pattern:

• LSTM: forward (with/without bias, multiple shapes), output subsets (Y only), bidirectional, reverse, initial h/c states
• GRU: forward (with/without bias, multiple shapes), output subsets, bidirectional, reverse
• RNN: forward (with/without bias, multiple shapes), output subsets, bidirectional, reverse

GRU round-trip tests (onnx → torch → onnx → ORT) use atol_onnx_torch2onnx=1e-4 because the manual loop exports as individual
elementwise ops rather than a single GRU node, causing minor floating-point order differences.

I was able to successfully load this CRNN model into PyTorch after these changes. I can not promise that it will work for every RNN model out-of-the-box.

Known limitations (documented in operators.md)

Operator	Unsupported
LSTM	layout=1, clip, input_forget=1, peephole weights, sequence_lens, dynamic weights
GRU	layout=1, clip, linear_before_reset=1, sequence_lens
RNN	layout=1, clip, sequence_lens, dynamic weights