Add authenticated AnyLearning remote inference

[vietanhdev]

Outcome

• adds an authenticated, bounded remote auto-labeling model backed by AnyLearning inference core
• keeps plaintext passwords out of YAML via environment-variable indirection
• requires verified HTTPS off loopback, rejects redirects and proxy inheritance, validates identities and response shapes, and cleans up token-owned jobs
• supports detection plus capability-driven point/box promptable models
• separates cancellation from unload so retrying does not destroy the client
• removes legacy process-wide TLS verification overrides
• adds a real authenticated TCP YOLOX-S E2E on Linux, Windows, and macOS with retained visual/results/timing artifacts

Local validation

• QT_QPA_PLATFORM=offscreen python -m unittest discover -s tests -v: 157 passed, 13 unrelated optional-model skips
• Ruff lint and format gates pass
• real official YOLOX-S ONNX, 3 runs: consistent bicycle/dog/truck results, no expectation failures
• retained and visually reviewed annotation: validation-results/Linux-local/20260830T165505Z-anylabeling-remote-c584467d/

Model provenance

The E2E pins the official YOLOX-S 0.1.1rc0 ONNX asset and dog image by SHA-256. Source and artifact are Apache-2.0; provenance is included in each retained summary.