Issue #783: python3 compatible - remove unnecessary imports.

A few improvements to function()

This commit is the first step of the plan to make Theano imports compatible with 3.x:

1. Remove unnecessary imports.

2. Convert relative imports to absolute.

3. Fix circular dependency issues by either eliminating "from xyz" imports or moving them inside functions.

Changes to make Theano work on Windows with Anaconda

Infer shape broadcast fix

Fix crash introduced in gh-1207 with theano variable that don't have the broadcastable attribute like sparse and random variable.

This is just confusing as it does not actually modify the function
being computed, as mentioned on the mailing list.

If the dictionary has <= 1 element in it, then we do not care if it's
unsorted.

Since they can be dictionaries, their emptiness shouldn't be checked by
comparing them to a list.

Fixed typo in doc of categorical cross-entropy

The true and coding distributions had been switched.
Also: a few PEP8 improvements.

CGemv now uses vector-vector dot product when appropriate

Doc

Infer shape broadcast

Needed for the ShapeFeature in later commit.

This solves a problem uncovered by the previous commit: a variable with
an ndim field could be replaced by a Variable without one.

dimensions that are broadcastable.

Also, a small modif to make those assert always work. Add a test for this.

This caused a crahs in the optimizer, as in one case, we didn't know
the shape was broadcastable by that path.

Cast indices to intp rather than intc.

The intent was correctly described in the comment - to cast to int32 on 32-bit machines and to int64
on 64-bit.  However, numpy.intc is not correct type to achieve that because it has the size of C int and
is 32-bit always.  The correct type is numpy.intp which is defined as "large enough to fit a pointer."

See http://docs.scipy.org/doc/numpy/reference/arrays.scalars.html .