Use IR in the optimizer

[justinchuby]

And improve robustness so it is reliable enough to be turned on by default.

[justinchuby]

The ExternalTensor support in the IR is working only when code is launched at the same directory as the model file. Aka. we read the external data file using relative paths. This is because when a model is loaded as a proto, its path information is lost to the IR. We need to design a way (e.g. pass in base path when deserializing) to handle when the model file is somewhere else.

This is relevant because I assume the optimizer will sometimes fold external tensors. Some examples may include Transpose(x) and element wise computation, casting, etc.

[gramalingam]

Related: #1436