run-command: add an asynchronous parallel child processor

run-command: add an asynchronous parallel child processor

This allows to run external commands in parallel with ordered output
on stderr.

If we run external commands in parallel we cannot pipe the output directly
to the our stdout/err as it would mix up. So each process's output will
flow through a pipe, which we buffer. One subprocess can be directly
piped to out stdout/err for a low latency feedback to the user.

Example:
Let's assume we have 5 submodules A,B,C,D,E and each fetch takes a
different amount of time as the different submodules vary in size, then
the output of fetches in sequential order might look like this:

 time -->
 output: |---A---| |-B-| |-------C-------| |-D-| |-E-|

When we schedule these submodules into maximal two parallel processes,
a schedule and sample output over time may look like this:

process 1: |---A---| |-D-| |-E-|

process 2: |-B-| |-------C-------|

output:    |---A---|B|---C-------|DE

So A will be perceived as it would run normally in the single child
version. As B has finished by the time A is done, we can dump its whole
progress buffer on stderr, such that it looks like it finished in no
time. Once that is done, C is determined to be the visible child and
its progress will be reported in real time.

So this way of output is really good for human consumption, as it only
changes the timing, not the actual output.

For machine consumption the output needs to be prepared in the tasks,
by either having a prefix per line or per block to indicate whose tasks
output is displayed, because the output order may not follow the
original sequential ordering:

 |----A----| |--B--| |-C-|

will be scheduled to be all parallel:

process 1: |----A----|
process 2: |--B--|
process 3: |-C-|
output:    |----A----|CB

This happens because C finished before B did, so it will be queued for
output before B.

To detect when a child has finished executing, we check interleaved
with other actions (such as checking the liveliness of children or
starting new processes) whether the stderr pipe still exists. Once a
child closed its stderr stream, we assume it is terminating very soon,
and use `finish_command()` from the single external process execution
interface to collect the exit status.

By maintaining the strong assumption of stderr being open until the
very end of a child process, we can avoid other hassle such as an
implementation using `waitpid(-1)`, which is not implemented in Windows.

Signed-off-by: Stefan Beller <sbeller@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>