llama: refactor llama_decode_impl

[JohannesGaessler]

This PR refactors llama_decode_impl by moving some of the code to functions llama_prepare_sbatch and llama_prepare_ubatch. There should be no change to functionality. The motivation for the change is that this enables re-using the code for training in #10544 .

[ggerganov]

In the #11213 refactoring, the llama_prepare_sbatch and llama_prepare_ubatch will be replaced with something like:

// internal llama_context logic will be implemented
// for example, the llama_prepare_sbatch logic will go here
llama_batch_manager_i bm = lctx.prepare_batch(batch, logits_all);

while (!bm->done()) {
    // the llama_prepare_ubatch() will be implemented here
    llama_ubatch ubatch = bm->next();

    ...
}

Feel free to merge. After merging, I will rebase the #11213 PR to illustrate this.