[opt-dist] Profile-guided optimization of LLVM regressed on Windows hosts

It has been found in https://github.com/rust-lang/rust/pull/143898#issuecomment-3068037546 that PGO build of LLVM is slower than general one a bit (~1%). Later in https://github.com/rust-lang/rust/pull/144255#issuecomment-3097557694 it has been proved that performance of MSVC dist compiler is better (the same 1%) (all in case of serde and ruff compilation). Zulip discussion has been with a question what to do with dist-msvc at the moment.
If you want to test it locally, here are the SHAs of above PR's jobs:
#144255: parent - 67819923ac8ea353aaa775303f4c3aacbf41d010; build - 978fa844ca3da87b313190c4b047f53a3cdb652f
#143898: parent - cccf075eba88363269e8589ebb8d40874cc542d8; build - a094db85f18d96658ed83010f55e8a24a68a26a0
for testing you may use rustup-toolchain-install-master tool (cargo install --git https://github.com/kennytm/rustup-toolchain-install-master), then install CI toolchains (rustup-toolchain-install-master --component=cargo <SHA> -i x86_64-pc-windows-msvc so you'll have cargo too) and then compile your favourite crate/program with hyperfine to get perf results (hyperfine "cargo +<SHA_PARENT> build --release" "cargo +<SHA_BUILD> build --release" --prepare "cargo clean" --runs 5, this will build code 5 times for each command to get a more reliable output)

please leave a cc for Windows folks and other who can help/do extra profiling, feel free to edit my message to add more info

Start with the performance comparisons in PRs #143898 and #144255 and the linked Zulip discussion. Reproduce the Windows MSVC results using the provided parent and build SHAs, rustup-toolchain-install-master, and hyperfine; done means identifying the cause of the regression and determining what should happen with dist-msvc PGO.