Switch to CompilerCaching.jl. - #3185

The callee walk in `typeinf!` re-infers callees via `typeinf_code` on the
same interpreter that ran the root inference cascade. The interpreter-local
cache can still hold tombstoned (LimitedAccuracy) const-prop results from
that deeper context, suppressing const-prop that would succeed in the
standalone entry and storing under-optimized source (e.g. unfolded Symbol
dispatch, failing GPU IR validation; JuliaGPU/CUDA.jl#3185). Clear the cache
before each re-inference, matching the runtime's fresh-interpreter-per-entry
behavior.

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Julia object references (julia.constgv globals) are kept session-portable by
stripping their initializers after codegen, with relocate_gvs! baking in the
absolute addresses at the toplevel link step. Leaving them behind as internal
globals with a null initializer made them fair game for optimization passes
that run before relocation -- notably the GlobalOpt run in the PTX back-end's
finish_module! -- which folded loads of them to inline nulls and deleted the
globals, constant-folding e.g. device-side comparisons against Symbol literals
(JuliaGPU/CUDA.jl#3185). Demote them to external declarations instead, which
no pass can fold, and have relocate_gvs! restore initializer and linkage.

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>