fix(llvm): Correctly cache pointers to intrinsic memory operations

[xdoardo]

Fixes the high CPU usage when using LLVM for modules that use atomics.

We were saving pointers to VM functions that operate on memory (for example, memory_notify and memory_wait) in the same cache. This led, then, to IR being generated like this:

if_then:                                          ; preds = %aligned_access_continue_block7
  %call_memory_notify = call i32 %xyz(ptr %0, i32 0, i32 %2, i32 1)
  br label %if_end

if_end:                                           ; preds = %aligned_access_continue_block7, %if_then
  %call_memory_wait32 = call i32 %xyz(ptr %0, i32 0, i32 %1, i32 %atomic_load, i64 -1)

where %xyz is a local variable the holds the value of the same function pointer resolved from cache, depending which between memory_notify and memory_wait, in this case, is added to the cache first.

[promptless]

✅ No documentation updates required.