View reuses the same internal tensors

[fmassa]

Currently, every forward/backward creates a new tensor (which shares the same storage).
This PR allows to keep track of the module's output/gradInput, as the tensor pointers won't change at every forward/backward pass.

I think that this behaviour should be enforced for all nn modules. If you agree, I can send a PR fixing this behaviour in other modules.

[soumith]

Totally agree. Have been writing new modules with this in mind.

[fmassa]

Great ! :) I'll have a pass over nn modules and I'll prepare a PR later today (or when I find some time)

[nagadomi]

This commit breaks older saved models. It does not have self.output.

In 14 module of nn.Sequential:
/home/nagadomi/torch/install/share/lua/5.1/nn/View.lua:80: attempt to index field 'output' (a nil value)

[fmassa]

@nagadomi sorry for breaking your loading. I haven't thought about the case in which one saves a model before ever forwarding/backwarding it. This should be fixed with #713.