Make concurrent retrieval configurable

[dionhaefner]

Currently, we use a multiprocess pool to retrieve RGB bands in parallel, and fall back to sequential retrieval if that is not available. I would prefer having this as a setting, potentially also with the possibility to use a thread pool. We moved away from threading after experiencing race conditions within GDAL, but that might not always be a problem (for example, the issue might be fixed in recent GDAL versions or only crop up in Lambda deployments). Since threading is often preferable for this workload it should be possible to use at the user's own risk.