Tests used to fail due to accumulation of negligible errors, so the
differences were big after a few iterations. Instead, we compare the results
at the end of each step, and synchronize the parameter values so that both
implementation restart from the same point.
Also, tightened the relative error bound.

reenable the automatic detection of when to don't load the full kernel for the gpu conv conv_patch_stack_reduce. This is just a speed up reenabled.

updated Doc and the handling of theano flags to don't pass him one linker.

lift subtensor taked on elemwise in more case. Test it. Don't lift if the elemwise is used by other computation.

lower a little bit theano overhead by removing call to enumerate as is it slow. Anyway, the index was already available otherwise.