speculative : fix batch sizes at initialization

[ggerganov]

ref #9949 (comment)

Fix batch size initialization to use the actual context of the model. Resolves issues when params.n_ctx == 0 (i.e. uses the training size of the model context).

[slaren]

I think this should fix the immediate issue, but there may be other issues. The batch should never exceed n_batch, which is always equal or smaller than n_ctx, so it should probably be initialized to n_batch instead, but that may indicate that the speculative example may create batches too big.
I also noticed that without -n, it would only generate one token, which probably indicates that it does not support the default n_predict of -1.

[ggerganov]

Indeed the example will currently assert with very large prompts that exceed the specified batch size: src/llama.cpp:17136: GGML_ASSERT(n_tokens_all <= cparams.n_batch) failed. llama-cli submits the prompt in a loop of batches so it does not have this problem.

We can leave it as it is though, since llama-speculative mainly focuses on the speculative functionality and don't think it is worth adding the extra logic for chunking the prompt into batches.