Add benches for ZeroMap

[sffc]

Follow-up from #652

Action items:

• Review how ZeroVec and ZeroMap work
• Learn Criterion: https://bheisler.github.io/criterion.rs/book/index.html
• Review the current benchmarks for ZeroVec: https://github.com/unicode-org/icu4x/tree/main/utils/zerovec/benches
• Add benchmarks for ZeroMap in Criterion
• Add similar benchmarks to LiteMap
• Optional: Add benchmarks for ZeroMap/LiteMap in IAI

@andrewnguyen

[Manishearth]

We should probably write the same benches for litemap too.

[Manishearth]

@andrewnguyen is interested in working on this

[Manishearth]

(Andrew isn't planning on fixing this anymore)

[Manishearth]

Alan plans to work on this

[poulsbo]

It's not clear to me what these benches are doing. I've put together a basic ZeroMap benchmark, and I can add more, but it would help to know our aim here.

I see with ZeroVec we quote benchmark numbers against Vec<T>, proving the speed of building/deserializing the thing, with comparable speeds for operations such as binary search, as expected. Do we want something similar here? We're binary searching the keys; do we want some numbers showing "serialization is a win, but you pay at lookup time"?

[Manishearth]

@poulsbo it's worth comparing against HashMap and also potentially LiteMap. We absolutely expect that lookup is more expensive, the question to answer is "how much more expensive" :)

utils/litemap already has a couple benchmarks to copy.

The main operations we care about are read operations, but benchmarking some write operations can be good too. Note that borrowed ZeroMaps will switch internal implementations if you try writing to them, it would be interesting to capture the performance impact of that.

[sffc]

it would help to know our aim here.

I would like to have results to allow us to expand this table to include ZeroMap.