Improve async load for adapters to avoid main thread lockups in server

[tgaddair]

The concurrency currently assumes that host execution time that holds the GIL is minimal, but particularly for loading adapters from disk to host memory, we see that large adapters can lock up the server.

This could in part be addressed by #434 which goes direct to GPU from disk with minimal host memory copying. We could also explore options like CUDA streams and pinned memory if needed to further optimize the load.