Fix align

Issue #783: python3 compatible - remove relative imports.

In python 3, list comprehention variables do not leak in the outside scope.  In this
case the loop is actually cleaner.

 - iter is a builtin
 - io is a standard lib module in 3.x.

Floor division operator for complex numbers is deprecated since python 2.3
and results in an error in 3.x.

With this commit theano/tensor/tests/test_basic.py passes under python 3.2.

 - disabled fix_filter becaus theano uses "filter" as a method name;
 - fixed another instance of checking the error message;
 - fixed another instance affected by numpy isinstance(x, int) bug.

  * allow numpy array as an argument to hash_from_code.
  * use truediv instead of div under 3.x.
  * **kwds checks that keys are strings in 3.x
  * fixed a typo bug in tensor/opt.py
  * fixed an incorrect numpy version check in test_basic

  * Use PyCapsule instead of PyCObject.

  * Use Python 3 module initialization.

  * Prepend 'm' to code hash to avoid module names that start with a digit.

Lazy config default value

  * Suppress 2to3's "next" fixer that conflicts with having non-iterator attributes called "next."
     Rely on six.next instead.

   * added python 3 version of hash_from_code.

Fix a test when g++ is not available.

Add triangle/nonzero functions

Try to fix the problem that depending of the inputs strides, the outputs strides of dimensions of size 1 get strides that make it considered not aligned by the PyArray_ALIGNED macro.