- 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

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

remove useless DebugMode warning when we think we where not returning a view of the input on empty ndarray.

* number fo threads on GPU must be a multiple of NUM_VECTOR_OP_THREADS_PER_BLOCK
* this means we will have more threads on the GPU than the number of streams we
actually want.
* the fix involves making these extra threads do nothing, instead of corrupting
the stream of random samples !

TODO: review MRG_RandomStreams test suite.

implementations (on CPU). DebugMode already takes care of this.

random numbers properly. This was hiding a GPU bug. Fix is in progress.

changed an assert to an Exception so that if the optimization fails it doesn't bring down the whole process