Add --target-sanitize=(memory|thread|address)

[xal-0]

Adds three hidden options that can enable the Memory/Thread/AddressSanitizer
LLVM passes even when Julia is not built with any sanitizers enabled. In
conjunction with #59033 this lets us build a ThreadSanitizer version of Julia in
only a few minutes rather than hours.

• Adds the --target-sanitize=(memory|thread|address) flags to enable those
  LLVM passes when compiling code ahead of time with --output-*.
• JITted code will enable the passes for a sanitizer only if the runtime is
  built with it.
• Adds the corresponding options to CodegenParams.

[Keno]

Can we spell this as --target-sanitize= just to follow the convention of other compilers?

[xal-0]

I'm also open to suggestions on ways to clean up the duplication between jl_options_t/jl_cgparams_t/OptimizationsOptions. I expect we'll accumulate more codegen knobs and AOT-only options because of juliac.

[Keno]

I think this is fine for now - we can think about refactoring the options struct at a later time.

[vchuravy]

Since these change codegen behavior, we should probably also thread them through the cache options (e.g.

julia/src/staticdata_utils.c

Lines 401 to 411 in 3b29187

	JL_DLLEXPORT uint8_t jl_cache_flags(void)
	{
	// OOICCDDP
	uint8_t flags = 0;
	flags |= (jl_options.use_pkgimages & 1); // 0-bit
	flags |= (jl_options.debug_level & 3) << DEBUG_LEVEL; // 1-2 bit
	flags |= (jl_options.check_bounds & 3) << 3; // 3-4 bit
	flags |= (jl_options.can_inline & 1) << 5; // 5-bit
	flags |= (jl_options.opt_level & 3) << OPT_LEVEL; // 6-7 bit
	return flags;
	}

& co) such that the native code in pkgimages is adjusted, and you can do tsan/asan with respect to it.

[Keno]

& co) such that the native code in pkgimages is adjusted, and you can do tsan/asan with respect to it.

We should also make sure that the sanitizer abi tags are still working and probably enable more of them on Yggdrasil

[xal-0]

I've added the sanitizer flags to Base.CacheFlags (by extending the UInt8 to a UInt64, so we hopefully don't have to do this again in the near future). Ideally this would use the same variable-length format as the new CPU features struct, but I don't think it's worth making that change now.

[xal-0]

Since the Base.CacheFlags change breaks Pkg.test(), this will have to wait for JuliaLang/Pkg.jl#4347.

[vtjnash]

Options that affect only the base image (--ouput-incremental=false) and get copied from there should not be put into the Base.CacheFlags struct, which is for flags that can be different from the base image.