5055:Port ProfileMode GPU to new back-end and profiling.py

Added meaningful message when missing inputs to scan.

Test abstract conv

Change grad to L_op 

Adding scipy.special modified Bessel function ops

Documentation on gpu float16 ops.

libraries not found by g++ are now printed to the user.

Diag --> AllocDiag

Prepare release 0.9.0rc4.

Various small fixes.

Make DebugMode keep memory layout when possible.

This will still convert a noncontiguous array to contiguous, but will
not convert F to C.  We can make it keep the noncontiguous layout but
it's more involved and can take a *LOT* more of memory.

fix gh-5694 I didn't test it, but it should now skip empty directory …

Fix offset in GpuIncSubtensor

docs: Fix openmp config section

Removed useless printf.

Clarify that the stable version of 0.8.2 and people don't want to use…

Fix error in Advanced indexing with > 2D