Compile PyTorch, torchvision, onnxruntime from source for numpy 2.x support

- PyTorch 2.8.0 with Jetson Orin optimizations (arch 8.7, ARM+CUDA linker optimization)
- Disabled unnecessary features (NCCL, QNNPACK, XNNPACK, FBGEMM, Kineto, etc.)
- torchvision 0.23.0 with CUDA support
- onnxruntime 1.20.0 with TensorRT EP
- flash-attn 2.8.3 (latest version)

Performance: 65.7 FPS (vs 62.2 FPS baseline = 5.6% faster)
Image size: 6.74GB (vs 8.28GB baseline = 18.6% smaller)

Size optimizations:
- cuDNN/TensorRT symlink preservation: ~2GB saved
- Remove test directories, dev tools, examples: ~500MB saved
- Conservative cleanup preserving public APIs (numpy.testing, torch.testing)

TensorRT optimization:
- FP16 precision enabled
- Engine caching enabled with 2GB workspace
- Builder optimization level 3
- Aux streams optimized for memory efficiency

Author: alexnorell

.github/workflows/docker.jetson.6.2.0.yml

@@ -19,8 +19,9 @@ env:
 jobs:
   docker:
     runs-on:
-      labels: ubuntu-latest
-    timeout-minutes: 120
+      labels: depot-ubuntu-24.04-4
+      group: public-depot
+    timeout-minutes: 360
     permissions:
       id-token: write
       contents: read