In the past it was warned when we load theano.sandbox.cuda.

Consecutive calls of the the join op ( during the optimization phase) get
the axis as a constant. What used to happen was that in such cases Theano
would not use that constant to define its broadcastable pattern.

Added .c files to setup.py to properly install the new lazylinker_c.c file -- also removed tabs in setup.py

The problem was that the version was obtained *after* importing the C module. As a result,
recompilation would fail because the library was still in use.
This change also saves the version in the __init__.py file of the lazylinker_ext base module,
so that it can be checked without loading the C module.

- can update shared variables internally, so that it does not need
  Function.__call__ to give correct answers.

- it can loop without returning to Python to implement scan (though this is not
  much faster than a for i in xrange(...) f.fn() type thing

- it is versioned now so that it is recompiled when necessary

- the scan test speed is modified to show these things.