[WIP] Implement cuDNN RNN ops and RNN module.

[dan-zheng]

This PR is a WIP, created to show progress and receive feedback. Please don't merge yet.

---

• [x] Implement and test wrappers for cudnnRNNForwardInference and cudnnRNNForwardTraining.
• [x] Implement and test cuDNN RNN module.
  • The Rnn module is designed to match PyTorch's recurrent layers, e.g. nn.RNN and nn.LSTM. It facilitates PyTorch-style model building.
  • Extract individual parameter tensors from single weight buffer.
  • cuDNN requires that all weights are contiguously allocated and passed as a single buffer. However, extracting individual parameter tensors is important for gradient descent, etc.
• [ ] Implement and test cuDNN RNN backward functions.
  • [x] cudnnRNNBackwardData seems to work.
  • [ ] cudnnRNNBackwardWeights needs debugging, the values are incorrect.
  • [ ] Update Module.registerParameters to correctly register arrays of parameters (ArrayBuffer[TensorR]).

[dan-zheng]

Reference PyTorch RNN program:

import torch
import torch.nn as nn
import torch.nn.init as init
import torch.nn.functional as F

input_size = 10
hidden_size = 40
num_layers = 2
seq_length = 5
batch_size = 3
bidirectional = False
num_directions = 2 if bidirectional else 1

rnn = nn.RNN(input_size, hidden_size, num_layers, nonlinearity='relu', bias=True, dropout=0, bidirectional=bidirectional)
for p in rnn.parameters():
    init.constant_(p, 0.01)
input = torch.ones(seq_length, batch_size, input_size)
input.requires_grad = True
h0 = torch.ones(num_layers * num_directions, batch_size, hidden_size)
output, hn = rnn(input, h0)
# Lantern produces the same output value.

output.backward(torch.ones_like(output))
print(input.grad)
# Lantern produces the same input gradient value.
for p in rnn.parameters():
    print(p.grad)
    # Lantern doesn't produce the same parameter gradient values yet.