Remove `IndexMut` impl from various collections to prepare for possible `IndexSet` in the future

[nikomatsakis]

I believe we intend to eventually extend the index traits with the ability to override the x[y] = z syntax specifically using a trait like IndexSet. This would allow hashmaps to be used like map[key] = value and have that insert the key, unlike today where map[key] = value will panic if key is not already present.

To pave the way, the simplest thing is to remove the IndexMut operator from those collections where we would anticipate growing on demand.

I believe that would be btree, hash_map, and possible vec_deque and vec_map.

Thoughts?

cc @aturon @alexcrichton @gankro

[nikomatsakis]

triage: P-backcompat-libs I-needs-decision (1.0 beta)

[nikomatsakis]

triage: P-backcompat-libs (1.0 beta)

[aturon]

I was just starting to prep a PR to resolve this by removing the impl, as you suggested. :)

[alexcrichton]

Don't we want to keep the implementation on HashMap for example to do something like let a = &mut map[key]?

[aturon]

@alexcrichton Depending on the exact way that IndexSet is done, it may not be possible to provide both modes of use. That is, we want to be able to write map[key] = value -- which we cannot today -- but it's not clear that you can do that and also support examples like the one you gave. In particular, that involves adjusting inference to determine whether or not a deref is in play.

For now, removing these impls will give us the most freedom when we do incorporate IndexSet.

[Gankra]

It seems weird to have this and insert -- insert has a return value, though. This would be a super-annoying ergonomic regression in the interim. What's the expected timeline for something like IndexSet? 1.1?

[Gankra]

Oh actually upon reflection most usages I know are Index, not IndexMut. Not as awful a regression.

[mahkoh]

@aturon

Depending on the exact way that IndexSet is done, it may not be possible to provide both modes of use.

It should be possible to move in, out, and take references with the [] operator. If that's not possible with the current design then the current design has to be changed instead of committing to yet another subpar design until 2.0.