fix: Update auto-selection to match MCP server defaults

Changed auto-selection in New Evaluation screen to match MCP server:
- API models (litellm/inference): cpu-basic (was: cpu)
- Local models (transformers): a10g-small (was: gpu_a100)

This ensures consistent defaults between UI and MCP cost estimator.

app.py

@@ -2757,9 +2757,9 @@ No historical data available for **{model}**.
             # litellm and inference are for API models → CPU
             # transformers is for local models → GPU
             if provider in ["litellm", "inference"]:
-                return gr.update(value="cpu")
+                return gr.update(value="cpu-basic")
             elif provider == "transformers":
-                return gr.update(value="gpu_a100")
+                return gr.update(value="a10g-small")
             else:
                 return gr.update(value="auto")