test new solver on CI until stabilization

[lqd]

We need the new solver to be able to build the stdlib, and this has regressed in the recent past. Things are broken right now, and we want to ensure this keeps working until stabilization later this year.

I've added this to the x86_64-gnu-llvm-XX job (note no -1/2/3 suffix) so it's run on PR and auto CI, and because it's roughly the fastest of these builders, though maybe the aarch64 llvm 1 and 2 are slightly faster.

Building the stage 1 ./x build library/ is generally quite quick, a couple minutes at most.

The new solver fix to build std is from @lcnr on zulip.
r? @marcoieni

[lqd]

Things are currently broken, the job should fail if this is working.

[lqd]

Fixing this comment made me realize running the mir-opt tests and --pass=check is duplicated in this job and in the llvm-2 jobs here

rust/src/ci/docker/scripts/x86_64-gnu-llvm2.sh

Lines 9 to 21 in 0e5924a

	# Run the `mir-opt` tests again but this time for a 32-bit target.
	# This enforces that tests using `// EMIT_MIR_FOR_EACH_BIT_WIDTH` have
	# both 32-bit and 64-bit outputs updated by the PR author, before
	# the PR is approved and tested for merging.
	# It will also detect tests lacking `// EMIT_MIR_FOR_EACH_BIT_WIDTH`,
	# despite having different output on 32-bit vs 64-bit targets.
	../x --stage 2 test tests/mir-opt --host='' --target=i686-unknown-linux-gnu
	# Run the UI test suite again, but in `--pass=check` mode
	#
	# This is intended to make sure that both `--pass=check` continues to
	# work.
	../x.ps1 --stage 2 test tests/ui --pass=check --host='' --target=i686-unknown-linux-gnu

So we could slightly optimize either job by removing these two tasks. I could also do that in another PR if you want?

View changes since the review

[rustbot]

This PR was rebased onto a different main commit. Here's a range-diff highlighting what actually changed.

Rebasing is a normal part of keeping PRs up to date, so no action is needed—this note is just to help reviewers.

[lqd]

Rebased to fix conflict