unrolled LSTM layer with batch BPTT

[xiw9]

An fast unrolled LSTM layer similar to BVLC/caffe#1873
Need 2 inputs,
The input sequence nodes_in[0] and the corresponding sequence label nodes_in[1].
nodes_in[0] size: [batch_size][1][1][input_width]
nodes_in[1] size: [batch_size][1][1][1]
Example config file:

data = train
iter = csv
  filename = "...."
  has_header = 0
iter = attachtxt
  txtfilename = "...."
iter = threadbuffer
  buffer_size = 4
iter = end

eval = val
iter = csv
  filename = "...."
  has_header = 0
iter = attachtxt
  txtfilename = "...."
iter = threadbuffer
  buffer_size = 4
iter = end

extra_data_shape[0] = 1,1,1
extra_data_num = 1

netconfig=start
layer[in,in_1->2] = lstm:lstm1
  nhidden = 1024
  parallel_size = 8
layer[2,in_1->3] = lstm:lstm2
  nhidden = 512
  parallel_size = 8
layer[3->4] = fullc:fc1
  nhidden = 51
layer[4->4] = softmax:softmax1
netconfig=end

# evaluation metric
metric = error

max_round = 40
num_round = 40

# input shape not including batch
input_shape = 1,1,4096

batch_size = 512

[xiw9]

conflict between txt iter and csv iter

data = train
iter = csv
  filename = "............"
  has_header = 0
iter = attachtxt
  txtfilename = "............."
iter = threadbuffer
  buffer_size = 4
iter = end

[antinucleon]

Thanks for your PR. I am still working on general case of LSTM, which is able to run CNN + LSTM together. In that branch we have special text IO for sequence. However I still have some bug to fix. Do you have interest to work together on that branch?