Traits for multiscalar multiplication

[hdevalence]

Split from #125: without settling on the design of the precomputation API, I think it makes sense to refactor the multiscalar functions to be traits implemented on the point types. Some of this is already done in the precomputation branch and could be extracted.

One question that would be worth thinking about is whether we want to stay with static dispatch for the iterators or to use dynamic dispatch. The downside of static dispatch is that it bloats the code size (maybe this doesn't matter) and that it prevents the multiscalar trait from being object-safe (which would prevent making opaque precomputation structs).

[hdevalence]

This is closed by #144.