feat: potential SLA breach batch

[luthfifahlevi]

Batch and consolidate potential SLA breach alerts per team

Summary

• IdentifyPotentialSLABreach now accepts repeated project_names and label_groups in one request instead of requiring a fan-out of multiple separate per-project/per-label calls, each of which triggered its own alert — teams previously got fragmented, noisy notifications instead of one clear summary.
• Server-side logic is split into pure detection (computeBreaches) and batch consolidation (IdentifySLABreachesBatch), which aggregates every project/label-group combo into exactly one alert per team, structured as project → SLA group → target job → causing jobs, tolerant of individual combo failures.
• Alert payload/model reworked to carry this structure end-to-end (proto, internal PotentialSLABreachAttrs, AlertManager adapter, Siren template); per-group severity was considered but dropped after team discussion in favor of a single request-level severity, keeping the request/response shape simpler.
• Adds a full local end-to-end test harness (colima + local Kubernetes + real Airflow + Optimus built from source) and documents it in dev/README.md, including several pre-existing local dev-setup bugs found and fixed along the way (missing required server config causing crash loops, a hardcoded gRPC port mismatch, a plugin-discovery path bug, and an Airflow Helm chart version pin that had gone stale).

Test plan

• go build ./...
• go vet ./core/scheduler/... ./ext/notify/...
• Unit tests: core/scheduler/service, core/scheduler/handler/v1beta1, ext/notify/alertmanager
• Verified end-to-end against a from-scratch local Kubernetes + real Airflow + Optimus environment (5 real jobs, actual KubernetesPodOperator execution, real dependency chain) — confirmed correct multi-level lineage attribution and real alert delivery to Siren rendering through the updated template
• Staging validation (pending)

[ahmadnaufal]

few more comments, other than that LGTM