Small improvements to pycaffe

[longjon]

I haven't tested these exact commits yet, although I've made the same changes elsewhere in the past.

These three commits:

• fix issue Python wrapper: classifier.py does not initialize mean if no mean file is provided #671, by correctly defaulting mean to zero even if set_mean is never called,
• reorder exceptions to provide a more helpful message for non-4d blob input, and
• change checks on channel_order and input_scale so that pathological input behaves closer to expectation.

Expanding on the last point in particular:

• Right now setting input_scale to zero is the same as setting input_scale to one, due to the check if input_scale. Allowing the multiplication/division by zero properly breaks things or produces warnings.
• Right now setting channel_order to () or 0 is the same as setting it to the identity, due to the check if channel_order. Note that both of these are valid indices in numpy, although actually using them in this context is probably a mistake. Allowing the indexing will make such usages crash in most cases, instead of continuing as if the input were valid.

[shelhamer]

Thanks for the Python fixes Jon! I'll combine with improvements from #735 and release to master.