feat(train_board_ocr): replace --ckpt-dir with --model-name and add hand MAE

[tkgstrator]

What

• Replace --ckpt-dir with --model-name (+ --ckpt-root, default ./runs). Checkpoints now land at <ckpt-root>/<model-name>/. model_name defaults to board-ocr-<backbone>, matching the sweep naming.
• --resume latest no longer errors when latest.pt is missing; it falls through to a fresh run so sweep scripts can pass it unconditionally.
• Add hand/mae to metrics and surface board_acc, hand_acc, and hand_mae in train + val log lines. MAE treats the count head as ordinal, so "off by 1" reads very differently from "off by 9".
• Update train.sh, train_multi_gpu.sh, train_backbones.sh to pass --model-name.

Why

• The --ckpt-dir interface forced callers to reconstruct the same ./runs/board-ocr-<backbone> path everywhere. --model-name collapses that to a single identifier.
• The old exact-match hand accuracy hid the difference between "1 off" and "9 off" — MAE makes ordinal quality visible in the training log.

Version

Bumped to v0.3.0 (CLI flag rename is breaking; pre-1.0 rule → minor).

Test plan

• uv run python -m mito_train.training.train_board_ocr --help shows --model-name / --ckpt-root and no --ckpt-dir
• Smoke run writes to ./runs/board-ocr-<backbone>/latest.pt
• Log line shows board_acc, hand_acc, hand_mae
• --resume latest on a fresh --model-name starts from epoch 1 instead of erroring

🤖 Generated with Claude Code