[vm] Inlining heuristics should take into account instruction sizes

[rakudrama]

Not all instructions at the inlining phase result in the same number of bytes of code. Inlining heuristics should do a better job of predicting code size.

In particlar,

• [ ] Redefinitions reference existing values, so potentially generate no code.
• [x] GenericCheckBounds annotated with pragma('vm:unsafe:no-bounds-checks') generates no code.
• [ ] The length() input to GenericCheckBounds annotated with pragma('vm:unsafe:no-bounds-checks') often becomes dead after the check is removed.

[a-siva]

//cc @alexmarkov