perf: go direct to _map_partitions in functions that receive already flat deps

[lgray]

Justification: if the dependencies are already flat there's no need to unpack them.

[martindurant]

I suppose we need a dask version pin or wait for the Task stuff?

[lgray]

I thought of something even better for this.

I think I just want to change map_partitions to check for nested collections (compare traverse=True to traverse=False deps) and if they're the same just go directly to _map_partitions. It'll save a lot of wrapping.

[lgray]

cool old tests pass. I'll devise a test to make sure it is doing what I expect.

[lgray]

so this is ready to go except for the fact it uncovered a bug in fuse linear chains. it does not respect kwargs for the older dask versions.

[martindurant]

except for the fact it uncovered a bug in fuse linear chains

Make that a separate issue and merge this?

[lgray]

I'd have to back out the change that allows non-collection kwargs to pass through if I didn't fix that issue.

[lgray]

But I do agree it's probably better to factor.