Orchestrator: Initialize LabelResolverCache from blu files when in a new Composer Session

[taicchoumsft]

Orchestrator training (i.e. creating embeddings for utterances) takes a long time. To alleviate as much pain as possible, try to cache results between builds.

Describe the solution you'd like

We already cache build results between builds. We store an in-memory map of LabelResolvers so that incremental training is very quick (only different utterances will need to be recalculated).

But when we start a new Composer session, the cache is cold and recalculation happens again, even though the embeddings have essentially been cached in the .blu files.

This request is to warm up the cache with data from the .blu files, so that training Orchestrator in a new Composer session will be quick, so long as we've trained before.