Make Portilla-Simoncelli texture model work on arbitrarily-sized images

[billbrod]

Because of how the Portilla-Simoncelli model handles multiscale representation, the input image needs to be divisible by 2^n_scales. There are two issue:

• po.tools.signal.shrink, which currently throws an exception if you try to e.g., downsample an image of size (225, 225) by a factor of 2. If you do the obvious thing (just round), the shrunken image looks wrong. It's probably an off by one error, so tracking that down will probably work. This happens in _reconstruct_lowpass_at_each_scale.
• There's something similar with po.tools.signal.expand as used in _double_phase_pyr_coeffs: if you expand an image of (128, 128) to (256, 256) and compare it against the finer-scale image of size (255, 255), you get a size error. Probably fixed by just lopping off one side or the other, but worth being careful.