- fixes errors in conv_patch_stack_reduce when the entire kernel doesn't fit
  into shared memory

- lowers the shared memory requirement of conv_patch_stack_reduce

Only print it for the device in use, when it is really being used. For the moment, just continue to print to sys.stderr.

Only parse THEANO_FLAGS once. Warn when a value-less key is provided. Use warnings instead of print statements.

Fixes http://trac-hg.assembla.com/theano/ticket/565.

should help identify which ares of the graph to optimize for GPU.

Updated Windows installation docs with instructions about how to get rid of the error message about assert not being declared

from previous runs, which ScanGrad would end up using

not run during import.  Running this code during import messed up the
device-switching logic (made it impossible to switch from default device).