mlgo AOT: multi-model support for the same policy

Support for embedding multiple models for the same optimization, and selecting between them with a compile-time flag. For example, we have an arm 32 and an arm 64 size-inliner in chrome. To LLVM, it appears as if there's only one model and the MLInlineAdvisor code is unchanged - except for passing an additional feature to this model, indicating which sub-model to use.

The exact design (extra feature) is perhaps an artifact of how this was implemented in Tensorflow. With EmitC-based AOT, this can be perhaps implemented more elegantly.