performance with very large input files

[UnixJunkie]

It seems the parallelization is not super good.
Maybe we should not cut the input file into chunks.
Instead, we should provide readers which move into the file using an offset,
if that's possible.

[UnixJunkie]

in the parallel case (not distributed), read_some does not need to really read from the input file:
it can emit a read instruction to the worker process.
For example:

• read block at index i with size s
• read line i to j
• skip M records then read N records

[UnixJunkie]

on one computer: the demuxer should translate Bytes|Line|Line_sep|Red into blocks
and the worker should take charge of reading that block

[UnixJunkie]

this will prevent from having to create tmp files to hold chunks

[UnixJunkie]

needs a bench also, to check that this really accelerate things