speculative : fix handling of some input params (#9963)

* speculative : fix batch sizes at initialization

ggml-ci

* speculative : handle params.n_predict == -1

* speculative : limit batch size to llama_n_batch