Use type tagging for OpaqueWrappable type checks

[jasnell]

Avoids using dynamic_cast/RTTI by using unique type tags for each type T. Perf gains here are expected to be negligible in typical benchmarks as there are other bottlenecks drowning it out... however, it's still a good micro-optimization / low-hanging-fruit update that carries very little risk.

[kentonv]

Hmm. Playing with godbolt I think there's a simpler version of this optimization:

When T is declared final, the compiler can optimize dynamic_cast<T*> to simply check if the vtable is equal to T's vtable -- very similar to the optimization you've implemented here, but without the need to store an additional pointer.

So if we simply declare the two OpaqueWrappable implementations as final then I think that's all we need. (Note that OpaqueWrappable<T, true> will have to be modified to not inherit OpaqueWrappable<T, false>, but instead copy the contents, but that's fine.)

I think this approach is preferable as it is easier to understand and doesn't require storing extra tags.

[jasnell]

Oh, nifty. Yeah, that's better. Updated to take that approach!

[jasnell]

Switched the PR back to using the type tag approach. Since we build with thin-LTO and visibility=hidden flags, and since benchmarks and some tests are compiled as shared libraries rather than static, we end up hitting a known llvm bug (llvm/llvm-project#71196) when attempting to use the approach @kentonv suggests in #5781 (comment). The original type tag approach works as an alternative.