Update doc with instructions for using new gpu backend

Make recent scan opt change deterministic.

Pool shape

TensorInv op added, test for tensorinv added

libgpuarray api changes

Update doc conv flags

Fix gh-4534. If ds size for pool is equal to image size and ignore_border=False, use the correct shape.

Change name of assert_no_cpu_op in old back-end as well

Speed up GpuDnnSoftmaxGrad by using the best mapping.

add 'assert_no_cpu_op' as a tag instead

Finish the regression fix in gh-4533

Make SearchSortedOp always return int64 as I'm not able to always figure out what is the output dtype.

In particular, in Python 32bit, it don't return int64, but if I make
it return intN where N is the python bit version, it also don't work.

remove redundant container checking.

The container is checked not None for twice

Better error related to gh-3376

Fix gh-4530, we can't have 2 opt with the same tag

Port local_assert_no_cpu_op to the new back-end

Add an include of gpuarray/buffer.h for future work.