Layer variable names about lbann

[timmoon10]

I propose the following name replacements for the following matrices in lbann_layer.hpp:

Matrix	Description	Proposed Name
WB	Weight-bias matrix	m_weights
WB_D	Gradient w.r.t. weight-bias matrix	m_weights_gradient
Zs	Output of forward pass linear transform	m_preactivations
Acts	Layer activations	m_activations
Ds	Error signal from "previous" layer	m_prev_error_signal
Ds_temp	Error signal to "next" layer	m_error_signal

[timmoon10]

Honestly, I think we should split up fully-connected layers, activation layers, and dropout layers so that we don't have to deal with the awkward Zs/m_preactivations matrix. In my proposed scheme, every layer will have four matrices: m_prev_activations, m_activations, m_prev_error_signal, and m_error_signal. m_prev_activations and m_prev_error_signal will point to the data in the "previous" layer if the data distribution is the same and will make a copy otherwise (thus, they will usually take up little memory and will save us the cost of doing multiple data redistributions). The current organization of lbann_layer.hpp requires that we pass an optimizer as an input, which limits us to optimize just one matrix. I suggest that we pass an optimizer factory to the layer and instantiate the optimizer internally. This will be useful if we want multiple optimizers (e.g. for filters and biases in convolutional layers). Finally, I don't like having an extra row of ones on the activations (so every activation matrix has NumNeurons+1 rows), but I haven't thought of a clean way to deal with this.