Better world age handling

[maleadt]

There's currently two related problems with our current world age handling:

• we index the compilation cache directly with the world age (by taking the hash of the compiler job); instead we should intersect, and handle invalidation
• when creating the FunctionSpec, we should have a way to pass the world age of the calling code, as opposed to the current global world age, which makes e.g. cufunction always look up the latest version of a potentially shadowed kernel (whereas it should be consistent with the world age of the caller); this probably requires a new intrinsic as both the current world age getter and the PTLS entry always point to the latest world:

   julia> worlds() = (ptls=Int(unsafe_load(convert(Ptr{Csize_t}, Core.getptls() + 8))), runtime=Int(Base.get_world_counter()))
   worlds (generic function with 1 method)
   julia> worlds()
   (ptls = 29611, runtime = 29611)
   julia> bar() = 42
   bar (generic function with 1 method)
   julia> worlds()
   (ptls = 29612, runtime = 29612)
   julia> Int(first(methods(worlds, Tuple{})).primary_world)
   29611