I am not sure why this is necessary now.  One guess is that the canonicalizer or
something else has been modified to put in identity(), and now it is more
obvious that we need an optimization to take it out again.  I'm hoping that this
improves the slowdown recently observed in convolutional net benchmark.

The modification ensures that that even if numpy is clever enough to not make a
copy on unit-length reductions, the perform() returns a copy.

Allow the default value of parameter dx,dy of ConvOp.__init__. They are set explicitly to 1 by the conv() fct.

remove FAST_RUN_NOGC from the acceptable value of flag mode. People should use the flag linker=c|py_nogc

In a test, test that we generate an error only when one is needed and test the type of the error generated.

In sparse.CSMProperties, don't put the last return as view, as we create a new ndarray from a tuple.

make debugmode use the new theano.misc.may_share_memory.may_share_memory fct to don't raise useless warning. It was calling the wrong *Type.may_share_memory sometimes.

added new theano.misc.may_share_memory that handle more case and can be called in all case(with and without scipy,...). Test it.