Support Ollama use: chat|memory and parallel embed beside VL.

Pull nomic-embed-text for Assistent memory, write ollama-roles.json, CPU Modelfile, and raise MAX_LOADED_MODELS/NUM_PARALLEL to 2.

Co-authored-by: Cursor <cursoragent@cursor.com>

tests/test_perf_tiers.py

@@ -34,8 +34,8 @@ def test_ollama_high_reserves_vram_for_swarm():
     )
     tune = ollama_tune_for(info)
     assert tune.flash_attention
-    assert tune.num_parallel == 1
-    assert tune.max_loaded_models == 1
+    assert tune.num_parallel == 2
+    assert tune.max_loaded_models == 2
     assert tune.kv_cache_type == "q8_0"
     assert tune.gpu_overhead_bytes == 14 * 1024**3
     env = "\n".join(ollama_env_lines(tune))