Llama improvements

[laggui]

Now that Llama 3.2 was added in this PR, we are missing the following improvements:

• Async checks (check stop criterion for stop token in background thread)
• Tensor cache should be fixed size (slice assign, not concat)

We are missing some ops for Top-P sampling, but we can have the first release with greedy sampling (argmax).

For Top-P sampling, we need:

• sorting (tensor.sort_descending_with_indices runs the default impl on CPU)
• cumsum (Burn PR missing cubecl kernel)