tag/untag box/unbox only on demand

[samrushing]

Something I want to try, especially for integers: do not enforce tagging in registers, but only in tuples. In other words, apply tag/box only upon store/fetch to/from a tuple. This would protect GC while theoretically really cutting down on tag/untag ops in code. While clang/gcc usually 'see through' the tag ops, it'd be cleaner and possibly faster. This would also have (good) implications for FFI.