Hc patchfrom part2

[Cyan4973]

just checking CI

[Cyan4973]

The new MT strategy in this PR favors keeping the jobs busy,
while the previous approach favored keeping I/O maximally completed.

As a consequence, this has an impact on the state of internal buffers at wake-up events,
and the --adapt mode, which was depending on them, doesn't work anymore.

This leaves a few option:

• Do not change the MT mode
• Do no longer support --adapt.
• Refactor the MT mode to make it compatible with --adapt
• Refactor --adapt to no longer depend on internal buffer state.

The last 2 items obviously cost a lot more time than the first 2 ones,
and this unfortunately may be the most critical ingredient in this equation.

[Cyan4973]

Changed for a combined strategy of waking up on flush event + waking up on job slot availability.
This requires sharing a condition variable, which is more complex to setup.

[Cyan4973]

This PR won't be merged. It's too complex and not stable enough.

I'll see if some of the simpler ideas can provide benefits and create a separate PR with them if that's the case.