Decision: Use of `asyncmap` in pmap batch mode

pmap in batch mode uses a local asyncmap to process a batch - https://github.com/JuliaLang/julia/blob/9e3318c9840e7a9e387582ba861408cefe5a4f75/base/distributed/pmap.jl#L198

Considering that each computation in pmap is fairly large, and batch sizes small, an asyncmap would not have a major overhead and if the computation involves IO, quite beneficial.

For example, if the input is a list of file names to be processed, it is efficient to interleave I/O and computation and hence a local asyncmap is a better fit.

This issue is to take a decision whether to

1. Keep it as it is, i.e., no change - the batch is processed using asyncmap

2. Change it to a local map. If the computation involves I/O the caller would have to explicitly partition the input and the mapping function in turn would need to perform an asyncmap and a flatten on the final output.

3. Add another keyword arg to pmap, batch_function=map. To use asyncmap, the caller would need to explicitly specify batch_function=asyncmap. A user defined function (for example one that uses @threads ) can also be specified.

Start with the batch-mode implementation in base/distributed/pmap.jl around line 198 and review how its local asyncmap is used. Compare the three alternatives in the issue, including their I/O and caller implications. Done means a choice is documented clearly, with any resulting pmap change identified.