Remaining scan errors

It seems that even the last fix of the generated random numbers didn't do
the trick. The reason being same numeric instability.
I've change the operation to a sum, which, if you apply it recursevly is
less problematic then multiplying numbers

Fixed crash at exit time if cache was deleted

Fixed some tests broken due to exception verbosity

This commit also includes a few PEP8 fixes.

This test fails because of small numbers (the gradients are especially
susceptible to small numbers). Increasing the range from which we sample,
it doesn't mean that there will not be a seed there for which the sampled
numbers are small enough (or a few entries are small enough) to get in that
instability region

This tests are of the sandbox of scan that I want to change anyway..

Gpu properties

Improved scan error message

I think that we can't really use iter(x) as suggested by David, because we want to iterate over the items.

Two improvements:
- An explicit error is raised if the lambda expression used in scan
  returns something that is not made of Theano variables (which may
  happen for instance when someone returns a constant value).
- An error is always raised when failing to parse the return value of
  the lambda expression

Fixed gh-356: dtype of tensor.sum()

Fixed some references to code recently moved

This is to accomodate changes from f9ca8f9d.

Gradient code is moved from tensor/tensor_grad.py to theano/gradient.py.
This makes it work with sparse variables.

This commit was originally written by Arnaud Bergeron. I re-authored it
to avoid a big merge in repo history.

Code for sparse was stolen from Yann Dauphin.

Make mlp_test test a more sensible case.

Previous test was testing settings where we removed either the
"local_track_shape_i" or "local_shape_to_shape_i" optimizations.
These are unlikely settings, what we test now is the complete
removal of the "ShapeOpt" optimization.

Fixed grad of grad of scan

Fixed tests with script run_tests_in_batch.py