Decouple from TensorFlow backend

[iamohcy]

Current code depends on multiple changes we have made to the TensorFlow codebase. This means that subsequent updates to TensorFlow require a long merge process to incorporate all the new changes, and ensure that they do not result in any changes that affect our implementation. Additionally, many changes are expected to the XLA codebase with the migration towards OpenXLA.

Goal is to reduce the changes to a patch that makes core functions that we modify virtual, so we can create a “plugin” style subclass that modifies the class as necessary. Specifically, implement a “HloEnvComputation” that subclasses the original “HloComputation” class with the new RewritePlan mechanism, we minimize the code changes necessary since we mostly just need to overload the ReplaceUseWith function.