Update test_hessian_bug_grad_grad_two_scans to run faster

Speed up scan test

Make sure to not delete empty folder when we do not have the lock

AdvancedSubtensor1 support for local_useless_subtensor optimization

Also check_py_code = False

Add a way for almost-equal types replacements to be patched over automatically.

Fix the problem with partial set_subtensor coverage in gpuarray.

Added GpuDownsampleFactorMaxGradGrad

Faster tests in DebugMode and more resistent to rounding difference.

This is a guess to try to fix strange return value for gpu_int_ptr

Feature for no outputs from inplace.