Support minchunksize

[MasonProtter]

I think a nice additional option would be a minchunksize that goes with nchunks. Basically, the way I'd want it to work is

scheduler=StaticScheduler(nchunks=10, minchunksize=5)
tforeach(f, 1:3; scheduler)        # does not spawn tasks (serial fallback)
tforeach(f, 1:10; scheduler)       # spawns 2 tasks
tforeach(f, 1:20; scheduler)       # spawns 4 tasks
tforeach(f, 1:100; scheduler)      # spawns 10 tasks
tforeach(f, 1:10_000; scheduler)   # still spawns 10 tasks

spawn 10 tasks. I think this is really useful for the majority of naively parallel operations on well behaved data when the mapping / reducing functions are cheap, because you can basically say "I want this parallelized, but don't create more tasks than necessary"

[carstenbauer]

Sounds like a good idea to me.

[carstenbauer]

Might make sense to support this upstream in ChunkSplitters.jl (cc @lmiq).

[carstenbauer]

Should be easily doable now that we use ChunkSplitters v3.x, which has this as minsize.