Fix parameter count

[luisherrmann]

This contribution introduces three mayor changes:

1. Parameters are now counted using the parameters() function of nn.Module and the numel() function of the nn.parameters.Parameter. This allows for counting the weights of custom modules with other keywords than just weight or bias. The fix addresses issues #95 and #65 .
2. Trainable parameters are now counted on a separate accumulator using the requires_grad attribute of parameters.
3. The API is expanded by adding a parameter recurse both to summary() and summary_string(), which is True by default.
   If set, the summary string will only be built for direct children of the model (and the model module itself).

The API extension and the parameter counting fix have both been tested through additional tests that are now included in torchsummary_test.py.