TensorRT: write image_size to onnx.yaml; fix demo output dir and intrinsics

[kvnptl]

What

• make_onnx.py: When exporting, write image_size: [height, width] into onnx.yaml using existing --height / --width, so the TensorRT demo knows the engine resolution without guessing.
• run_demo_tensorrt.py: Only rm -rf out_dir when it is not the same as onnx_dir and onnx_dir is not under out_dir — avoids deleting .engine files when both point at e.g. output/. Always os.makedirs(out_dir, exist_ok=True).
• run_demo_tensorrt.py: Scale intrinsics after resize with K[:2] *= …[:, np.newaxis] so NumPy broadcasts correctly after fx / fy.

Why

• Missing image_size in onnx.yaml caused wrong resize vs engine shape and reshape errors.
• Unconditional rm -rf out_dir deleted engine artifacts when out_dir and onnx_dir overlapped.
• K[:2] *= [fx, fy] failed with a broadcast error for (2,3) vs (2,).

How to verify

1. Run make_onnx.py with --height / --width; confirm onnx.yaml contains image_size.
2. Run run_demo_tensorrt.py with --onnx_dir output/ and --out_dir output/ — engines should remain; run should complete through point cloud.