Use device configuration from models.yaml instead of hardcoding to cpu

[rogermt]

Problem

Device preference is hardcoded to "cpu" throughout the request pipeline, ignoring the device: "cuda" setting in models.yaml.

Current Flow

1. Server defaults to "cpu" (api.py L124)
2. Server does NOT include device in tool_args (tasks.py L388-391)
3. Plugin falls back to "cpu" when device not in args (plugin.py run_tool)
4. models.yaml device: "cuda" is read by load_model_config() but never used in request path

Expected Behavior

• Load device config from models.yaml
• Pass device to plugin via tool_args
• Plugin uses configured device instead of defaulting to cpu

Related Issue

• #164 (Phase 12: plugin argument contract)

Files Affected

• server/app/api.py - Device parameter handling
• server/app/tasks.py - tool_args construction
• plugins/forgesyte-yolo-tracker/configs/models.yaml - Device config
• plugins/forgesyte-yolo-tracker/src/forgesyte_yolo_tracker/plugin.py - Device usage