feat(eval): add classifier support

[Stephen Belanger (Qard)]

Summary

Ports the classifiers feature from the Ruby SDK (braintrustdata/braintrust-sdk-ruby#154) into the Go SDK, implementing the cross-SDK classifiers spec.

• New `eval.Classifier[I, R]` interface and `Classification` struct mirroring `Scorer`/`Score`. Returns categorical labels (`id` + optional `label` + optional `metadata`) rather than a numeric score; multi-label classifiers may return several `Classification` items.
• Each classifier runs in its own span with `type: "classifier"` / `purpose: "scorer"` and input/output JSON matching the spec wire format. Per-case results are aggregated onto the eval span as `braintrust.classifications`.
• Classifier failures are non-fatal: per-classifier errors are aggregated and merged into the eval span's `braintrust.metadata.classifier_errors` map. Failing classifiers do not abort other classifiers or scorers.
• Classifiers run concurrently with each other (one goroutine per classifier) and concurrently with the scorer pass within each case.
• Adds a runtime validation that at least one of `Scorers` or `Classifiers` must be provided (spec MUST).
• Adds a customer-facing example at `examples/classifiers/` and a kitchen-sink internal example at `examples/internal/classifiers/`.

Test plan

• `make lint` — 0 issues
• `go test ./eval/... -count=1` — all tests pass, including 9 new classifier tests
• `go test ./eval/ -race` — clean under the race detector
• `go build ./...` in the examples module — both new examples build
• End-to-end smoke against a real workspace: `go run examples/classifiers/classifiers.go` and `go run examples/internal/classifiers/main.go` — verify classifier spans appear with `type=classifier` and aggregated classifications render in the UI

🤖 Generated with Claude Code