Should we fall back to serial execution for nchunks=1 and chunksize>length(input)?

[carstenbauer]

In both cases, we end up with a single chunk and already now run effectively serially. However, the difference to SerialScheduler is that we

• run a different code path and thus
• still create and spawn a (single) task which might run on a non-main thread.

[arnaud-ma]

Maybe having a new argument like single_chunk_mode = :serial or a boolean one serial1chunk = true can be a good idea. I think this is related to #114. One of the reasons for having minchunksize is to have (almost) no overhead when minchunksize > length(input).