Pooling classes can't return output dimension with unspecified image size

[rightaditya]

The Pooling classes should be able to provide output dimensions when the image size is unspecified for use with ConvolutionalSequence, but this currently raises an exception:

>>> from blocks.bricks.conv import MaxPooling
>>> m = MaxPooling((1, 1))
>>> m.get_dim('output')
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python3.5/dist-packages/blocks/bricks/conv.py", line 289, in get_dim
    ignore_border=self.ignore_border, padding=self.padding))
  File "/usr/local/lib/python3.5/dist-packages/theano/tensor/signal/pool.py", line 189, in out_shape
    r += padding[0] * 2
TypeError: unsupported operand type(s) for +=: 'NoneType' and 'int'

This is a relatively simple fix---the Pooling.get_dim() could just do a check---but it looks like it was written with the same idea as Convolutional---that None could be handled by just passing it down to Theano. Unfortunately it looks like the Pool.out_shape() in Theano doesn't like that. I've opened this issue to see if others would prefer trying to rely on Theano to make the change---if just doing a check in Pooling.get_dim() makes more sense, I'd be happy to submit a PR for it or include it in #1047.

[dwf]

It'd be nice if Theano supported that case but I think it's fine to special-case our code. Feel free to submit a PR.