perf: NATS-level baseline throughput/latency with `nats bench`

[e7217]

Context

EDG advertises "ultra-low latency" and high-performance ingestion (README), but no published baseline numbers exist for the NATS data plane. Before we can reason about regressions, capacity planning, or the impact of validation overhead in internal/core/handler.go, we need a documented, reproducible baseline at the NATS layer — independent of EDG's processing pipeline.

This issue covers the NATS-level baseline only. End-to-end load through the EDG core (handler → validation → JetStream → Telegraf) is tracked separately in the scenario-based load test issue.

Goals

1. Reproducible script at scripts/perf/nats-bench.sh wrapping nats bench for:
   • Core NATS publish throughput (platform.data.asset subject)
   • JetStream publish throughput (PLATFORM_DATA stream, matching cmd/core/main.go config)
   • Pub/Sub round-trip latency (p50, p95, p99)
2. Documented baseline at docs/perf/baseline-nats.md:
   • Hardware/OS context (CPU, memory, disk type, NATS version)
   • Result table for varying message sizes (e.g., 256B, 1KB, 4KB) and publisher counts (1, 8, 32)
   • Distinction between core NATS and JetStream numbers (the gap is the durability cost)
3. README/docs link so future perf work can reference the baseline.

Out of Scope

• EDG core handler/validation overhead (covered by scenario load test issue)
• Long-running soak tests (>10 min)
• Multi-node NATS clustering
• CI integration — this is run-on-demand for now

Acceptance Criteria

• scripts/perf/nats-bench.sh runs against a local NATS+JetStream and prints structured results
• docs/perf/baseline-nats.md includes at least one full result matrix with environment metadata
• Stream config used in benchmark matches cmd/core/main.go:79-95 (FileStorage, retention, replicas)
• README points to the baseline doc from the performance/reliability section

Notes

• nats bench ships with the official nats CLI — no new build dependency.
• Pair with ADR 0001 (reliability: Specify JetStream/ack semantics and add chaos tests #65) so JetStream numbers reflect the documented reliability config.