Smart context handling

[AlexCheema]

Perhaps something like https://github.com/tinygrad/tinygrad/blob/master/examples/llama3.py -- this doesn't prefill part of the prompt that's already been filled, it's super simple to implement.