Add batched_generate_fn()

[apaz-cli]

Still a bit rough around the edges. Debug prints are commented out. But it does the thing, so putting it up for review and feedback. Writing tests and cleaning it up.

• Does handle:
  • Stop sequences (supports all litgpt models)
• Does not yet handle:
  • Batched input_pos
  • Padding to the left
  • include_eos

Once include_eos works, I'd like to rewrite generate_fn() in terms of batched_generate_fn().

Also we're going to have to re-benchmark everything of course.