Graphgen improvements

[fmassa]

Should fix https://github.com/fmassa/optimize-net/issues/6 There is a lot of magic going on in here:

• temporarily overwrite torch __index and select to keep track of the modules in which those functions were called (thanks upvalues !)
• if the input tensors are not found in nodes table, look for them in the trickyNodes table, which contains the tensors created via __index and select during the forward call
• additionally, remove the need of having to run forward before generating the graph to pre-allocate self.output, by adding the output edges after each module's call.

cc @szagoruyko