Don't double-count inference time

[pchintalapudi]

Fixes #48024

During #46825, the timing code was updated to use task-local state instead of unsynchronized globals. Unfortunately, one of the pieces of timing code ended up using a different state tracker than the other timers, which led to the overcounting in #48024. That is fixed here + a test.

[IanButterworth]

Nice! The added tests are great, but I wonder whether we should also explicitly test the string output of @time given that @time comes with particular expectations for how code is compiled that might change if the macro changes. So parse the first percentage in the string etc.

Something like this perhaps

fname = tempname()
open(fname, "w") do
    redirect_stdout(f) do
        @time foo()
    end
end
buf = read(fname)
matches = collect(eachmatch(r"(\d+(\.\d+)?%)", buf))
@test parse(Float64, split(matches[1].match, "%")[1]) <= 100 # comp
@test parse(Float64, split(matches[2].match, "%")[1]) <= 100 # recomp