Selective operator build workflow for multiple models in a single runtime environment

🚀 The feature, motivation and pitch

ExecuTorch provides several mechanisms for selective operator compilation (e.g. EXECUTORCH_SELECT_OPS_MODEL, EXECUTORCH_SELECT_OPS_YAML). These workflows appear to be primarily designed around building a runtime for a single model.

In deployments where a single runtime must support multiple models, each model may require a different set of operators. Currently this typically requires manually extracting and merging operator sets from multiple models before building.

It would be able to provide multiple models or YAML files as input for operator extraction, or to document a recommended workflow for building a runtime that supports multiple models.

This would simplify common scenarios where one ExecuTorch runtime needs to execute several models.

Alternatives

Have been using a solution by merging multiple YAML files outside of Executorch, so that is a usable workaround.

Additional context

No response

RFC (Optional)

No response

Start by reading the existing EXECUTORCH_SELECT_OPS_MODEL and EXECUTORCH_SELECT_OPS_YAML workflows and how operator sets are extracted. Compare their current single-model behavior with the reported external YAML-merging workaround; done should mean either multiple models or YAML files are supported directly, or a documented workflow clearly explains how to build one runtime for several models.