Add workload-level profiling (draft — provisory)

[artursarlo]

⚠️ Draft / provisory — not yet ready for review. Opening early to run more precise end-to-end tests against upstream before requesting review. Scope, test plan, and this description will be refined.

Brings Pinterest's workload-level profiling feature onto upstream as a single, self-consistent commit.

Backend

• Heartbeat workload inventory sync (diff/upsert instead of delete+insert) into structured inventory tables, avoiding sequence write amplification.
• GET /profiling/workload_status endpoint with server-side pagination, filtering, and a ~30s precomputed snapshot store for scale.
• Supporting dynamic-profiling utilities, metrics/filters models, and DB manager methods.

Frontend

• Profiling Status console view (workload-aware inventory) with filter deep-links, partial-match filtering, and pagination wired to the workload_status endpoint.

Schema & ops

• Postgres migrations for structured workload inventory, workload precompute store, and heartbeat sequence cache tuning; fresh-install schema updated to match.
• periodic_tasks refresh_workload_snapshot job to rebuild the snapshot.

Testing

• Local e2e harness and Kubernetes sandbox (Spark-on-k8s workloads, per-thread Java profiling) plus backend/integration/spec tests.

Backend requirements add bitmath; date-fns is already declared for the frontend locale helper pulled in transitively.

Verified: backend import graph internally consistent (py_compile clean, import closure converged); frontend import closure resolves.

Co-authored with Pinterest contributors (see commit trailers).