proto: Remove write_source (pull in `IndexMap` to avoid borrowing issues)

[gretchenfrage]

This PR is competing with and mutually exclusive with my previous #2242 and #2230. I think this is the best iteration so far.

Successfully removes WriteSource without utilizing a callback (per #2230) nor introducing lifetime stuff (per #2242), and successfully splits up the write_source method's logic more deeply than those two.

Introduces dependency on indexmap crate. This seems fine to me. It's an extremely common crate and likely to be useful for other things too.

Notes on possible follow-up work

As I've expressed on the other PRs, I like the idea of introducing some sort of public API capable of doing what WriteSource can do. This PR is actually more helpful to that goal than #2242. In #2242, introducing some sort of quinn version of WriteGuard<'_> would mean we'd want to create a struct which contains both a mutex guard and a proto::WriteGuard<'_> which borrows from that mutex guard. This probably would require us to either pull in the ouroboros crate or write equivalent unsafe code manually, or to complicate the API in some ugly way. Conversely, with the approach implemented in this PR, we could conceivably just make write_limit and write_unchecked public and give quinn a WriteGuard struct which contains a mutex guard and a usize write budget. Or something like that.

Also, a different bit of follow-up work that could be done after that would be to modify RecvStream to do a similar sort of preemptive hashmap lookup, if we wanted.