v0.11.0 -> v0.12.0: startup hardware self-assessment (Roadmap J)

- hardware.py: assess_hardware() probes RAM, CPU, GPU VRAM (rocm-smi),
  llama-server, Qdrant, SearXNG reachability — writes hardware_state.json
- routers/hardware.py: GET /api/hardware (no auth) returns snapshot
- app.py: calls assess_hardware() in lifespan after init_db()
- 4 new tests: all services reachable, rocm-smi absent, llama unreachable,
  HTTP endpoint

hardware_state.json

@@ -0,0 +1,16 @@
+{
+  "ram_total_gb": 15.2,
+  "ram_available_gb": 13.4,
+  "cpu_count": 16,
+  "vram_total_mb": 0,
+  "vram_free_mb": 0,
+  "llama_reachable": true,
+  "llama_models": [
+    "/home/gramps/models/Mistral-Nemo-Instruct-2407-Q4_K_M.gguf"
+  ],
+  "qdrant_reachable": true,
+  "qdrant_collections": [
+    "jarvis_rag"
+  ],
+  "searxng_reachable": false
+}