Changing `build.profiler` forces an unnecessary rebuild of std

[Zalathar]

In theory, opening config.toml and toggling build.profiler between false and true should be cheap. Setting it to true should cause the profiler runtime to be built, without rebuilding std. And setting it to false should simply cause the rest of the build to pretend that any previously-built compiler runtime doesn't exist.

In practice, toggling build.profiler does trigger a rebuild of std, which also naturally triggers a rebuild of the compiler.

I think this is an artifact of library/profiler_runtime being considered an optional dependency of std. And I think that is just an artifact of how the profiler runtime was originally added (#42433), not something that inherently has to be true.

Of course, the tricky part will be fixing this without accidentally breaking anything.

[Zalathar]

I think bootstrap itself is behaving correctly here, in that if the profiler runtime really were a dependency of std, then enabling it should trigger a std rebuild.

So it's more of a questionable configuration, in that it really shouldn't be a dependency.

Also, toggling the profiler off/on is presumably not very common, especially now that the profiler can be built without the src/llvm-project submodule (#129295). So this should be a pretty low priority relative to its difficulty.

[onur-ozkan]

Additional note: "profiler" feature can still be enabled on std through this option:

rust/config.example.toml

Line 770 in 798fb83

#std-features = ["panic_unwind"]

even if we remove it from the build.profiler logic.