Shape inference + deterministic graph evaluation

[Human9000-bit]

Some of the most obvious and straightforward method of onnx optimizer improvement is doing shape inference

But in order to deterministically measure impact of new optimizations, I think we need a some sort of model graph evaluation.
But graph evaluation could pretty much use shape inference. I am submitting 'em in a singular PR due to this.

The idea is some part of runtime executor can make better decisions with known shapes.

[Human9000-bit]

(if CI passes)
I think we can merge this PR as-is, as graph cost eval is mostly done. After merge, I can make another PR where I actually plug shape inference into the model loader itself