[Tingliangstu]

Summary

Adds torchnep compatibility to pesmaker plot train while preserving the existing GPUMD/NEP plotting path.

Changes

• Detect torchnep training output from output.log or the torchnep loss.out header.
• Parse torchnep loss.out files with a text header and plot the two training stages separately in the loss panel using solid and dashed segments, plus a Stage 2 marker.
• Report torchnep explicitly in the CLI summary, including the Stage 2 start epoch and total epochs.
• Filter tensor parity rows with missing labels (nan) so torchnep runs without virial/stress labels still produce valid energy and force plots.
• Allow --source to point at an output file path by resolving to its parent directory, and auto-detect an output/ directory candidate.
• Add regression coverage for two-stage torchnep outputs.

Validation

• python -m pytest tests\test_training_nep.py
• python -m pytest
• PYTHONPATH=src python -m pesmaker plot train --source \\wsl.localhost\Ubuntu-24.04\home\tingliang\software\my_GPUMD\GPUMD\torchnep\example\PbTe\output --output-dir %TEMP%\pesmaker_torchnep_plot_check