Emit acquire fence before loading final field

[armanbilge]

In typelevel/cats-effect#3899 (comment) @durban brought to my attention that the current implementation of final Field Semantics when loading a final field is not correct.

Specifically, we must issue an aquire fence prior to loading the final field. This prevents the load of the field from being re-ordered before the load of the object reference (e.g. due to a compiler optimization that attempts an early, speculative load). Such a re-ordering could cause a pre-initialized value for the field to be observed.

Unfortunately I'm still not confident that we will have a correct implementation of final Field Semantics after this change. However, I do believe that this change brings us closer to that.

On the other end of the spectrum, I am hopeful that we might eventually be able to remove this fence entirely. Ideally, the dependency of the load of the final field on the load of the object reference would enforce the ordering. I don't have sufficient imagination to see how the compiler could generate speculative code that violates this ordering. If we can be confident that neither LLVM nor supported hardware will perform such a re-ordering, then it could be safe to remove this fence.

[WojciechMazur]

Don't worry about any failure in the CI. These are related to the bugs in the GC and I'm trying to fix them