as well !)

resulting in "undefined variable" error.

Added unit test test_inc_subtensor.py (almost identical to test_incsubtensor) to
test the new (undeprecated) inc_subtensor / set_subtensor functions

Oops, bug fix in (and streamlined) error reporting when updating shared variable (implemented with James)

left-padded if the user asked NOT to store intermediate results (this should be
the default-case for shared variables)

updated documentation related to verify_grad and TensorType

Very confused why this optimization is necessary...

split the test test_GpuCrossentropySoftmax1HotWithBiasDx in two to don't use GpuDot22 in it to make it fast enought for older gpu card. So the test don't fail on older gpu card.

enabled test_mlp again as now it work correctly. Disable it when the optimization are not the normal one as it is too slow.

remove duplicate check that make exit of theano longer then necessary with compiledir with many compiled element.