Instead of suppressing calls to free after closing the cuda context, delay closing the cuda context until all calls to free have been made. Might not be *quite* as performant, but feels perhaps a little cleaner.

Track whether there is a gpu context active; if not, don't try to free memory. Note that trying to free the memory and then just not displaying the error doesn't work, since the act of trying to free the memory creates a new gpu context.

None is almost like False, except it is OK do downcast a Python float to
a scalar floatX array. Other specialized case might be added later.

Change all theanorc code blocks in docs to use the 'cfg' Pygments filter which is designed for this.

Implemented unittests which test that the contract for In is respected: we test
that aliasing the input is impossible with borrow=False.

Old theano versioning system is deprecated with DWF's latest changes. Versioning
is now handled through setup.py and mercurial hooks.

borrow argument for In, gradient of a mtrix ..)

are tested.

Fixes compatibility issues with MacHG and probably GUIs. (Note that the released version of MacHG actually doesn't work at the moment, but new ones should soon -- cf. http://bitbucket.org/jfh/machg/issue/141/p2d2-environment-variables-not-set)

Be a bit more liberal in what we accept for a join operation. No longer force all broadcastable flags for non-join axes to be the same, and the output's broadcastable flags are the OR of the input's for each axis.