Execute tasks for independent pipelines in parallel

[tomwhite]

Currently, when computing something like xp.add(a, b), a is computed first, followed by b, followed by the add operation.

We could get more parallelism by computing a and b at the same time, since they don't depend on each other.

(This is true for the Lithops and Modal DagExecutors which both rely on map_unordered, but not for BeamDagExecutor, which can already compute a and b in parallel since it delegates to a Beam DAG.)

[tomwhite]

For Modal we could use the same idea as AsyncPythonDagExecutor to merge async streams for parallelism.

Lithops isn't async so that won't work. One idea I had was to compose multiple map calls into one by tagging the function and the inputs so that each input is processed by the relevant function.

[tomwhite]

I've written an initial implementation for Modal here: https://github.com/tomwhite/cubed/tree/compute-arrays-in-parallel

[tomwhite]

Fixed in #259 and #263