Local fix for max_pool2d differentiation
A recent change made it so max_pool2d isn't replaced by
max_pool2d_with_indices on the JIT path. This means it cannot be
differentiated correctly since max_pool2d_with_indices_backward requires
indices. Fortunately, we don't need the indices, but this prevents it
from being landed to core PyTorch. This is a stop gap to get the tests
working again, it looks like core PyTorch should be fixed.
