This fixes a crash when running tests on Windows.

In the buildbot make opt error raise and not warn.

Correctly cast the scalar during opt with dot22scalar.

Fix for some shape-related errors on 32-bit

Added regression test for bug fixed in blas opt

edited NEWS.txt

Added new command 'theano-cache unlock'

This can be used to unlock the compilation directory without having to wait
two minutes, or having to manually delete a file.

This test ensures that integer scalars are properly cast to floats to be
included within a dot22scalar op.

Minor fix in an theano/tensor/blas.py.

First version of file for the release candidate.

Changed argmax dtype from int32 to int64

Refactored GPU install instructions

Also refactored a bit GPU tips, and removed the note about limitation to
4 GPUs (which is not true anymore).

Moved instructions from "Using the GPU" tutorial page to the intallation
instructions.

Added a link from installation instructions to the "Using the GPU"
tutorial.

Minor stuff

Fix argmax grad

Some test fixes

Fixed some tests with compute_test_value = ignore

Now also testing gradient with axis=None (failing as of this commit).

If 'axis' is empty then x.ndim must be equal to zero, and thus we also
have 'axis == range(x.ndim)'.
If x.ndim in non-zero, there is no reason for axis to be empty, so
better not catch this case.

Windows requires pickled files to be opened in binary mode.

Also rewrote file operations to be safer.

Fixed test_lazy.py and test_vm.py

Everything looks fine.