llama: use FA + max. GPU layers by default (#15434)

* llama: use max. GPU layers by default, auto -fa

* ggml-backend: abort instead of segfault

tools/server/tests/unit/test_speculative.py

@@ -14,6 +14,7 @@ def create_server():
     server.model_draft = download_file(MODEL_DRAFT_FILE_URL)
     server.draft_min = 4
     server.draft_max = 8
+    server.fa = "off"
 
 
 @pytest.fixture(autouse=True)