performance: add benchmarks #191

Motivation

We want every performance PR to show how much it helped. That's not currently possible because:

Current limitations

• The draft benchmarks in chore: add pytest benchmarks #165 cover only XarrayReader, on a small Zarr store with 4 time steps. They don't cover the mosaic backend or the /info endpoint, where fix: open each dataset once per request #171 and the info() slowness live.
• They have no Icechunk tests.
• They have no long time axis. TEMPO has 17,000+ time steps, and the info() problem only shows up with that many.

Propose solution

1. Count tests, with no new dependency. Ordinary pytest tests that count the expensive things a request does and assert a limit. For example:
   • "a tile request opens each dataset once"
   • "a warm tile request makes one store request"
   • "a tile request creates no boto3.Session"
2. Timing benchmarks (revive chore: add pytest benchmarks #165) for CPU-bound changes:
   • rebase it and drop dask (it was only there to compare against feat: filtering based on multiple variables #151);
   • add a long time-axis fixture (17,000+ steps; coordinates only, so it stays small);
   • add a local Icechunk fixture;
   • add app-level cases through TestClient (/tiles, /point, /info, /info?show_times=true).

Each PR compares before and after with pytest-benchmark compare.

CodSpeed could later run the same benchmarks and post a comparison on each PR, because pytest-codspeed accepts the same benchmark fixture. Its default mode measures CPU instructions, so request-bound improvements still need the count tests.

Acceptance criteria

• Counting helpers exist for dataset opens, store requests and boto3.Session constructions.
• There is at least one count test each for /tiles, /point and /info.
• chore: add pytest benchmarks #165's benchmarks run on main and include the long-time-axis, Icechunk and app-level cases.
• The benchmarks README explains how a PR reports its before/after.