[MangelMaxime]

Add documentation about DOTNET_TieredPGOported from #5135.

Based on experimentation, we can't make it the default because the improvement only works in single build mode and make the watch mode worth.

Watch is the daily driver so we don't want to make it worse.

Good candidates to set this env variable is CI or if you don't rely on watch mode (AI are able to use the watch mode for faster feedback loop).

Numbers

Numbers, Fable phase on tests/Js/Main (75 files), medians of interleaved runs:

Single build, --noCache:

Cores	PGO on	PGO off	Change
32	2.2 s	1.8 s	-18%
4	4.7 s	4.2 s	-11%

Watch mode, same project, 8 edits per scenario:

Scenario	PGO on	PGO off	Change
First compile of the session	2.20 s	1.75 s	-19%
Body edit (21 dependents re-checked)	428 ms	508 ms	+19%
Signature edit	865 ms	950 ms	+10%

The first compile saves about 400 ms and each rebuild then costs 70 to 100 ms more, so a watch session is net slower after four to six edits.

It also depends on the project. From the measurements in #5135 (compile phase, median of 5, PGO off vs on):

Project	PGO on	PGO off	Change
Fable JS test suite	3.83 s	3.19 s	-17%
Fable Python test suite	3.16 s	2.62 s	-17%
Sutil	4.88 s	4.63 s	-5%
Ionide	3.02 s	3.06 s	+1%
Fantomas tools	4.47 s	4.85 s	+9%

Note

Happy to make it the default in the future, if we guarantee that it is always a win. Another reason, to make it the default could also be if the watch compilation is so fast that 10-20% worse has an almost negligible impact.