Add manual system telemetry metrics logging

[bghira]

Summary

• add a training-side system telemetry sampler for trackers without native host metrics
• collect CPU, memory, output disk, network rate, and GPU metrics
• support NVIDIA/CUDA, AMD/ROCm, and Apple/MPS telemetry sources where available
• skip manual telemetry for WandB to avoid duplicating its native host metrics

Refs #3167.

Validation

• .venv/bin/python -m unittest tests.test_system_metrics -v -f
• .venv/bin/python -m py_compile simpletuner/helpers/training/system_metrics.py simpletuner/helpers/training/trainer.py
• git diff --check