Fix attention IO capture for keyword-only Llama modules

Fix dtype mismatch during wave tank replacement training

Use provided HF token explicitly for download and push

Update wave_tank_surgeon.py

Add wave tank surgeon script