Fix Pooling output dimensions w/ unspecified input dimensions

[rightaditya]

The Pooling classes break when trying to provide output dimensions when the input dimensions are unspecified. This is particularly important for use with ConvolutionalSequence. This fixes that with a simple check on the input dimensions rather than passing the Nones down to Theano's Pool.out_shape(), which won't handle them. Also included is a regression test, as well as an addition to the ConvolutionalSequence regression test.

Fixes #1048.