[GraphOptimizer] Fix bug in post-lowering optimizations for OpenCL

[shajrawi]

As part of #3452 I created a layout verifier for the OpenCL backend. It caught a bug in our optimizations: see smallConv backend correctness test with the OpenCL backend for a tiny repro. We have a NHWC conv. which we turn into NCHW with a transpose of the inputs (which is fine). then the bug occurs when we optimize post lowering: OptimizeTransposeIntoReshape turns the transpose (of a placeholder) into a reshape and DCE removes the reshape. this is obviously incorrect because the placeholder is in NHWC format.

[nadavrot]

Great catch @shajrawi !

[shajrawi]

Found out why DCE deletes the Reshape node, at some point the conv. is no longer its user: it is done in the OptimizeReshape optimization, we do Reshape(Constant) -> Constant'. after the placeholder has been turned into a constant in placeholder->constant optimization.
This is fine, the verifier complained because it needs to know the layout of any constants / placeholders. The original constant had NHWC layout, and if we try to find out the layout of Constant->Conv. using Glow's canonical layout requirements then the new layout is NHWC too. OptimizeRehsape is not aware that it is running on OpenCL backend and as such it created a constant with NHWC instead.