Make Rust benchmark builds faster

Benchmarks are very slow to build, mostly because (1) we have a very large library and (2) we have many benchmarks binaries, which each need to independently link to our very large binary. Unless I set --jobs 1, I run out of memory trying to build them.

For the microbenchmarks, most of our benchmarks are our optimized distance functions and linear algebra functions. If we put them in a separate sub-crate (lance-compute?), then it should be must faster to link the microbenchmarks. We might even be able to consolidate the binaries for an additional benefit.

For macro benchmarks, we should probably move them over to the Python benchmarks in the top-level benchmarks directory. Then we just have to link one Python library and we can re-use it for as many benchmarks as we want.

TODO

• #1218
• Separate algorithms into smaller library and benchmark separately.

Start by inspecting the existing Rust benchmark binaries and the top-level benchmarks directory, then measure which builds consume the most memory and time. Determine how the microbenchmark algorithms and macro benchmarks are currently organized; done means separating the algorithms into a smaller library and benchmarking them independently, with faster builds and reduced memory use.