Remove the build directory as soon as possible when installing a package

[kit-ty-kate]

Fixes #5884
Queued on #6912

[kit-ty-kate]

Tested over 40_000 times with great success together with #6440. Max disk + RAM usage goes from 4-10GB down to < 2GB per install.

[rjbou]

Génial! lgtm on the idea.
tiny comment + it worth adding a bench if possible.

[rjbou]

This comment is tied to which line ?

[kit-ty-kate]

the one above. See b6ed206

[kit-ty-kate]

tiny comment + it worth adding a bench if possible.

done. I've added the bench in #6912 and rebased this PR on top of it to check the result.

[kit-ty-kate]

Compared to baseline (#6912), the benchmark shows a 0.3% performance degradation of regular installs and a 5% performance improvement on IO-constrained filesystems (simulated as much as i could)

[rjbou]

We need a final check on the bench to rerun once rebased, and if it confirms it should be ok (there is no reason that the gain changes)

[rjbou]

Compared to baseline (#6912), the benchmark shows a 0.3% performance degradation of regular installs

What is the reason of this degradation ?

[kit-ty-kate]

What is the reason of this degradation ?

rm of the build dir are sequential and block the next install when installing fast packages to build so it's very slightly slower when the builds are too fast

[kit-ty-kate]

Ignoring the build failure on cygwin (#7030)

Compared to baseline (#6912), the benchmark shows a 0.3% performance degradation of regular installs and a 5% performance improvement on IO-constrained filesystems (simulated as much as i could)

Latest diff is -0.1% and -1.3%, well within the margin of fluctuation. The main improvement is RAM usage over time anyway, not the performance improvement.