Dask client should be optional

Currently, a dask client / cluster connection is needed for fitting.

However, it should be possible to not use this but instead do the fitting in a the same thread (blocking)

ThreadPoolExecutor is an option but was found to be very slow for PyTorch fitting

No file or test is named. Start by locating the fitting entry point and tracing where the Dask client is required, then assess how a blocking same-thread mode could coexist with the current execution options. Done means fitting works without a Dask connection and the behavior is covered by tests without regressing existing fitting.

Vom Indexierungsmodell aus dem Issue-Text verfasst.