Benchmarks and performance tuning

[sunshowers]

For the initial release I've focused on getting the API right much more than performance. Idiomatic Rust tends to be quite fast, but there's likely quite some space to optimize.

I'd start by writing some benchmarks for common operations (or maybe adapting publicly available benchmarks), then noticing if some operations are particularly slow. If you're interested, please post in this issue!

[sunshowers]

One note is that we currently use a hash map to map indexes to items. We could probably use a vector of maybe-uninitialized items + a bitset instead. (A bitset could also act as a fast free list if we want to try and be as compact as possible.) There's some opportunity to be clever here.