feat: add GRU layer alongside LSTM

[ulises-jeremias]

Only LSTM exists. GRU is a simpler and often more efficient recurrent unit.

Task

Following the pattern of nn/layers/lstm.v and nn/internal/lstm.v:

1. Add nn/layers/gru.v — GRU layer definition
2. Add nn/internal/gru.v — CPU forward/backward kernels
3. Add GRU gate files for autograd
4. Add GRU test case

Stretch

• CUDA/Vulkan forward if GPU support is desired

Verification

• GRU forward matches PyTorch (nn.GRU) for same weights
• Backward gradients are correct