Systematic performance measurement

We have been doing a lot of _ad hoc_ measurement of the performance and file size of FlatGFA, but we need this to be more systematic. Here is a list of desiderata:

* Compare FlatGFA's running time and file size against other tools, including odgi.
* For performance, measure the time it takes to convert files (from GFA to both `.og` and `.flatgfa` formats) and for several analysis operations (e.g., `fgfa depth` vs. `odgi depth`). Make it easy to add comparisons for new commands.
* Use a collection of medium- to large-sized GFA files as input. Make it easy to add new files.
* Automatically run the benchmark suite on every commit to `main` and on every PR. Possibly use an existing commercial product, such as [Codspeed](https://codspeed.io) or [Bencher](https://bencher.dev) for this.
* Also provide a convenient way to run the benchmarks manually & locally (i.e., it shouldn't be _necessary_ to push to `main` to run the benchmark suite).

I have started a very rough sketch of what this might look like [in `bench.py` in this repository](https://github.com/cucapra/pollen/blob/e0cf9284b411fe201e8168b8e4d978765e1cfc9c/bench/bench.py). This does not yet do close to everything we need, and we may need to throw it away and start over, but there are a _few_ things that I like about the current design. First, there is a configuration file for declaratively specifying which experiments to run, like this:
https://github.com/cucapra/pollen/blob/e0cf9284b411fe201e8168b8e4d978765e1cfc9c/bench/config.toml#L1-L20

The tool can also automatically download sets of graphs from public repositories, as configured here:
https://github.com/cucapra/pollen/blob/e0cf9284b411fe201e8168b8e4d978765e1cfc9c/bench/graphs.toml#L1-L4

That seems kind of nice. I am also happy with my little Python wrapper for [hyperfine](https://github.com/sharkdp/hyperfine) invocations:
https://github.com/cucapra/pollen/blob/e0cf9284b411fe201e8168b8e4d978765e1cfc9c/bench/bench.py#L68

It's a simple, reasonably safe interface to the underlying tool executions.

To make the evaluation harness of our dreams, I suggest we proceed with these steps:

* [ ] Make a truly MVP measurement tool that can only measure one thing: for example, GFA conversion time.
* [ ] Try to set up CI for this tool, hopefully by integrating with Bencher or Codspeed. For example, Bencher seems to support [custom benchmark results](https://bencher.dev/docs/how-to/track-custom-benchmarks/); we could report our results in this format or just use its existing harness off the shelf. Bencher even seems to have a mode for [reporting file sizes](https://bencher.dev/docs/how-to/track-file-size/), which is relevant to us.
* [ ] Expand our coverage of both graphs and measured commands.
* [ ] Write documentation for other team members, so they can easily add their command comparisons to the benchmarking workflow.