Fix different .cachepath initialization points for precompile load

[Keno]

For quite some time I've been observing Revise randomly not pick
up changes to my packages. Usually I just write that off to Revise
getting into a bad state and restarting Julia fixes it. Today, I
got very annoyed by this again and decided to finally track it down.
What turns out to happen here is the packages in question are those
which:

A. Are being precompiled on load
B. Use Requires
C. Have an @require'd-dependency already loaded
D. The change to be revised is made in a file not loaded via
Requires.jl.

In this case the __init__ of the package triggers Requires,
which in turn calls back to Revise, which tries to start watching
the package. However, on the compilecache path (but not on the
path where we just find a pre-existing cache file), we used to
not set the .cachepath property of pkgorigins until after
the __init__ callbacks run, causing Revise not to be able
to see those files. Infuriatingly, restarting julia fixes this
because it just loads the .ji file that was just compiled.

Fix this by unifying the point at which the .cachepath is set,
always setting it just prior to the module initialization callback.

[KristofferC]

Makes sense to me at least.

[vtjnash]

It feels like this also might need to be dealt with on the compiled=false path too? I tried to unify that somewhat already (#43257), but it went mediocre.

[Keno]

It feels like this also might need to be dealt with on the compiled=false path too?

The particular revise issue was specific to .cachepath, which is not required for the compile=false path. I agree that it would be better to be very clear about what is initialized when and to make sure it's consistent to the greatest extent possible, but that's probably a larger change.

[timholy]

Thanks for looking into this & fixing it!