Fix libnvdla_compiler.so COPY path for Jetson 6.x runtime

[alexnorell]

Summary

• Corrects COPY --from=builder source path from /usr/lib/aarch64-linux-gnu/tegra/libnvdla_compiler.so* to /usr/lib/aarch64-linux-gnu/nvidia/libnvdla_compiler.so* in both Dockerfile.onnx.jetson.6.0.0 and Dockerfile.onnx.jetson.6.2.0.
• The previous path didn't exist in nvcr.io/nvidia/l4t-jetpack:r36.x, so the glob silently matched zero files and produced an empty layer. That layer cached forever, so the fix in PR Fix missing libnvdla_compiler.so in Jetson 6.x TRT runtime #2201 never actually shipped a working libnvdla_compiler.so into the runtime image.

Why this matters

On Jetson 6.x devices, ONNX Runtime's TensorRT execution provider needs libnvdla_compiler.so. Without it:
```
[E:onnxruntime:Default] Failed to load library libonnxruntime_providers_tensorrt.so:
libnvdla_compiler.so: cannot open shared object file: No such file or directory
EP Error ... Falling back to ['CPUExecutionProvider'] and retrying.
```
The model then runs entirely on CPU, ~3x slower (we observed 21 FPS camera vs. 1.3 FPS inference on an AI1/Orin NX).

Verification

Inspected nvcr.io/nvidia/l4t-jetpack:r36.4.0 directly:
```
$ docker run --rm --platform linux/arm64 nvcr.io/nvidia/l4t-jetpack:r36.4.0
sh -c 'find / -name libnvdla*'
/usr/lib/aarch64-linux-gnu/nvidia/libnvdla_compiler.so ← actual location
```

Test plan

• Trigger Jetson 6.2.0 container build via workflow_dispatch with a custom tag and verify /usr/local/cuda/lib64/libnvdla_compiler.so* is present in the resulting image.
• Pull on a JP6.2 device, run inference container, confirm ORT loads TensorrtExecutionProvider instead of falling back to CPU.
• Compare inference FPS before/after on a representative workflow (expecting ~3x improvement).

[alexnorell]

Superseded by #2321