optimizer: use count checking framework

[ianatol]

In #44557, we fixed some places where SSAValue uses were miscounted. During the development of that PR, after an egregious amount of printlns, we devised a framework to test our calculated use counts against an "oracle" use count independently derived from the relevant IncrementalCompact.

This PR cleans that framework up a bit and adds an option to toggle it (its off by default of course). Hopefully this framework means that no one has to manually check these counts again if/when we decide to make related changes.

Turning this on during our compiler tests shows we do still overcount at times. Overcounting is a lot less bad than undercounting (which was the crux of an issue with #44557 where we were deleting actually used instructions), but it may mean that we are missing some opportunities to do dead code elimination.

For example, the following comes almost immediately after starting Base.runtests("compiler"):

julia> compact.result.inst
29-element Vector{Any}:
 :(Base.getfield(_3, 1, $(Expr(:boundscheck))))
 :(Base.getfield(_2, :itr))
 :(Base.bitcast(UInt64, 1))
 :(Base.sub_int(%3, 0x0000000000000001))
 :(Base.arraylen(%2))
 :(Base.sle_int(0, %5))
 :(Base.bitcast(UInt64, %5))
 :(Base.ult_int(%4, %7))
 :(Base.and_int(%6, %8))
 :(goto %3 if not %9)
 :(Base.arrayref(false, %2, 1))
 :(Base.add_int(1, 1))
 :(Core.tuple(%11, %12))
 :(goto %4)
 :(Base.nothing)
 :(goto %4)
 :(φ (%2 => false, %3 => true))
 :(φ (%2 => %11))
 :(φ (%2 => %12))
 :(φ (%3 => %15))
 :(goto %5)
 :(goto %7 if not %17)
 Core.PiNode(:(%20), Core.Const(nothing))
 :(return %23)
 :(Core.tuple(%1, %18))
 :(Base.add_int(%1, 1))
 :(Core.tuple(%26, %19))
 :(Core.tuple(%25, %27))
 :(return %28)

julia> compact.used_ssas[13]
1

julia> oracle_used_ssas[13]
0

In this case, our oracle shows that %13 = :(Core.tuple(%11, %12)) is unused (which we can visually confirm), but it has not been eliminated by DCE because we think it has a use somewhere.

[Keno]

Concept LGTM. One DRY comment on the implementation.

[ianatol]

Gonna add some basic userefs tests just to make sure there's nothing else along those lines here

[ianatol]

@Keno I think this is ready now --- the interface to actually use the checking framework is a bit clunky, but I'm not sure of a better way to throw such an error from inside Core.Compiler.

Eventually, once we've gotten SSA counting to a place where the oracle check consistently passes, I'd like to add some tests that utilize `oracle_check