[ATfL][build.sh] Help CMake to pass the COMPILER_RT_TARGET_HAS_ATOMICS test

[pawosm-arm]

The product_build/runtimes/runtimes-bins/CMakeCache.txt file generated
by CMake insists that there are no atomics in the compiler-rt library:
//Test COMPILER_RT_TARGET_HAS_ATOMICS
COMPILER_RT_TARGET_HAS_ATOMICS:INTERNAL=

This is due to a failing test which does not actually link the compiler-rt
library, which results in a linker error and failed CMake test:

    ld.lld: error: undefined symbol: __aarch64_cas8_acq_rel
    >>> referenced by src.cxx
    >>>               CMakeFiles/cmTC_7f91d.dir/src.cxx.o:(main)
    ld.lld: error: undefined symbol: __aarch64_ldadd8_acq_rel
    >>> referenced by src.cxx
    >>>               CMakeFiles/cmTC_7f91d.dir/src.cxx.o:(main)
    >>> referenced by src.cxx
    >>>               CMakeFiles/cmTC_7f91d.dir/src.cxx.o:(main)
    clang++: error: linker command failed with exit code 1

See product_build/runtimes/runtimes-bins/CMakeFiles/CMakeConfigureLog.yaml
for a complete log.

This causes a cascade of subtle problems down the line, including
faulty PGO instrumentation code being injected into instrumented
binaries, which causes race condidtions whenever the instrumented
code is multi-threaded.

This patch helps linking the compiler-rt library when building the
runtimes.

Note that for this patch to have any effect, a companion upstream PR
needs to be merged which fixes a bug in the
compiler-rt/cmake/config-ix.cmake file: llvm/llvm-project#224681

[pawosm-arm]

Rebased and adjusted to the most recent changes.

[pawosm-arm]

I need to rework this patch. I should have spotted this sooner in the CMake files...:

# TODO(PR51389): We should check COMPILER_RT_BUILTINS_LIBRARY and report an
# error if the value is NOTFOUND rather than silently continuing...

This suggests which unexposed problem should be addressed first.

[pawosm-arm]

I need to rework this patch. I should have spotted this sooner in the CMake files...:

# TODO(PR51389): We should check COMPILER_RT_BUILTINS_LIBRARY and report an
# error if the value is NOTFOUND rather than silently continuing...

This suggests which unexposed problem should be addressed first.

That's not all, and that's not even a stopper. The real bug also needs to be fixed and I've positively tested a fix. I'll open an upstream PR soon(-ish).

[pawosm-arm]

The upstream part has been merged: llvm/llvm-project#224681