Add a few optimizations to reduce the number of copies

[udoprog]

This adds infrastructure to delay the assembly of values so that they can if appropriate target the stack directly instead of always having to resort to copy the value from a position in the stack to the top of it.

This has several benefits:

• This reduces peak memory use, since the stack will overall be smaller.
• This reduces the number of ref count increase/decreases a value has to go through when operating over it. Like we do with the IndexGet operation (see use of Stack::address_ref which can even use a reference of the value).

My own benchmarks shows a 50% reduction in the number of copies in aoc_2020_1a and a ~10% performance gain.

Because this has to live side-by-side with the existing assembler, Asm has to support both modes:

• Asm::apply will cause the assembly to generate the necessary instructions. No compiler state will be modified, since this is currently the norm.
• Asm::apply_targeted will produce an InstTarget appropriate for the operation. If a value is copied onto the stack, it will be declared as anonymous. This should be used immediately in a scope so the scope can be cleaned up.