Implement matrix sharding

[zanieb]

Since we've encountered the GitHub matrix size limit of 256 entries a couple times now (recently in #388), this adds sharding to our matrix generation script.

There's now a --max-shards option which splits the matrix into (up to) the requested number of shards. We don't split it evenly, we eagerly fill each shard first. To consume the shards, we just copy the build job into two identical child jobs which consume each matrix shard. Ideally, we'd avoid code duplication here by using a re-usable action or something for the build-steps — but that's overkill at this point.

We only need this for Linux for the foreseeable future.

[zanieb]

e.g., new warning and error

❯ uv run -q ci-matrix.py > matrix.json
warning: matrix of size 288 exceeds limit of 256 but sharding is not enabled; consider setting `--max-shards`
❯ uv run -q ci-matrix.py --max-shards=1 > matrix.json
error: matrix of size 288 requires 2 shards, but the maximum is 1; consider increasing `--max-shards`

[zanieb]

e.g., we create empty shards when empty

❯ uv run -q ci-matrix.py --labels arch:foo --platform darwin --max-shards=2 | jq
{
  "0": {
    "include": []
  },
  "1": {
    "include": []
  }
}