fix: harden AI Studio JSONL upload checks

[YOIMIYA66]

Summary

This PR turns the MedRecord-CN ERNIE SFT run findings into reusable safeguards for the cli-trainer skill.

• add a --upload-plain-file fallback that writes JSON/JSONL training files through the AI Studio Git contents API
• optionally remove .gitattributes JSON/JSONL LFS rules before uploading plain training files
• make --verify-upload --strict-lfs detect Git LFS pointer text, not only the is_lfs field
• document the root-filename workaround for trainDataFiles 指定的文件不存在
• parse PaddleFormers learning_rate: log lines so artifact export can produce LR charts

Why

During the MedRecord-CN full SFT run, the SDK/CLI upload path could leave a JSONL file as an LFS pointer or make the training mount fail even when the Git API could see the file. The reliable path was to remove JSON/JSONL LFS rules, upload the actual JSONL content as a plain file, verify pointer content and size, then submit with a short root filename such as train.jsonl.

Validation

• python -m py_compile scripts/train.py
• python scripts/train.py --help
• helper smoke check for LFS pointer detection and learning_rate metric parsing
• git diff --check HEAD~1 HEAD
• token scan for exposed AI Studio credentials