feat: CUDA backward for LSTM, Attention, BatchNorm, LayerNorm, Embedding, Pooling

[ulises-jeremias]

Only Linear and Conv2D have CUDA backward. All other layers lack GPU backward:
LSTM, Attention, BatchNorm, LayerNorm, Embedding, MaxPool, Dropout (training mode).
This blocks full GPU training for models using these layers.

Task

For each layer type, add CUDA backward kernels:

• nn/internal/lstm_backward_cuda_d_cuda.v
• nn/internal/attention_backward_cuda_d_cuda.v
• nn/internal/batchnorm_backward_cuda_d_cuda.v
• nn/internal/layer_norm_backward_cuda_d_cuda.v
• nn/internal/embedding_backward_cuda_d_cuda.v
• nn/internal/pool_backward_cuda_d_cuda.v

Verification

• Each backward matches CPU reference within 1e-4 relative error
• Gradient check utility confirms correctness