Build the stdlib from rust-src sources.

[eddyb]

See rust-lang/rust#69060 (comment) for background.

I definitely feels slower than previous. We might want to build libstd only once and place it in the sysroot (and I believe that would mean we can skip downloading rust-std altogether).

EDIT I moved from -Zbuild-std to a custom cargo build of the stdlib sources, no slowdown left.

However, we only need this for a couple weeks, until rust-lang/cargo#8073 (hopefully) reaches beta.
Then all of the hacks here will be unnecessary, since Rust itself will be built fully deterministically.

[eddyb]

rust-lang/cargo#8073 landed, so we shouldn't need the "fake version" trick anymore.
(well, once rust-lang/rust updates src/tools/cargo to include that)

[eddyb]

This took 5h09m to finish on a 24-core/48-thread server. Somehow I don't think we can do it this way :).

[eddyb]

Tried the same run but without -Zbuild-std, it finished in 1h20.
So this PR makes perf runs almost four times slower. That's unacceptable.

[eddyb]

1h23 with the new approach. Also, for fun, I looked inside the rlibs built here and saw:

 00000000: 7275 7374 0000 0005 0047 bd06 2b72 7573  rust.....G..+rus
 00000010: 7463 2031 2e34 342e 302d 6e69 6768 746c  tc 1.44.0-nightl
-00000020: 7920 2838 3335 3432 3863 3335 2032 3032  y (835428c35 202
+00000020: 7920 2864 6637 3638 6335 6338 2032 3032  y (df768c5c8 202
 00000030: 302d 3034 2d31 3529 0463 6f72 65da 81b4  0-04-15).core...

(this is expected, since rustc outputs that into the rmeta format to be able to detect incompatibilities)

What's nice is that there are no other differences remaining.

[Mark-Simulacrum]

Thanks! I want to take another look in the morning and merge and deploy then.

[eddyb]

So the tests I did locally are with a recent noop change for which we now have official results too.

For "syn-opt clean":

	Before	After	Δ%
perf.rust-lang.org	37,490,787,593	38,584,396,789	2.9170%
This PR (locally)	37,521,722,991	37,521,495,334	-0.0006%

So I believe this PR should remove the problem (but we should keep an eye on results for a while).

[eddyb]

Oddly enough, this had a beneficial perf effect.
EDIT: looks like it's entirely from shaving ~10ms off of run_linker with locally-built stdlib somehow.
EDIT2: just remembered our locally-built stdlib has no debuginfo, that might be it.