we just burned multiple hours on runpod setup failures for flux lora run 2. lock this down now.

fix /static/train_anky_setup.sh so it is resilient on fresh pods:

1) blackwell compatibility:
- default torch install to cu128 wheels instead of cu124
- keep it configurable via env if needed

2) token UX:
- if HF_TOKEN is missing, prompt interactively in the terminal
- if ANKY_TOKEN is missing, prompt optionally (allow empty for no live heartbeat)

3) dataset extraction robustness:
- our archive extracts to final-training-dataset-for-round-two, not training-images
- detect common extracted folder names and rename to /workspace/dataset automatically
- fail loudly if no dataset folder with pngs is found

4) venv robustness:
- if /workspace/venv exists but /workspace/venv/bin/activate is missing, rebuild it
- ensure python3-venv/python3-pip installed up front

5) docs:
- update /trainings/general-instructions so next time we remember the real one-liner, token behavior, cu128 requirement, and recovery steps

goal: one command on a new runpod should reliably get to tmux training launch without manual surgery.