tts : fix n_ubatch + make WavTokenizer cache-less

[ggerganov]

fix #13712

• Set the WavTokenizer n_ubatch == n_batch since it is computing embeddings
• Use llama_encode() instead of llama_decode() with the WavTokenizer context
• The WavTokenizer does not need a KV cache, so we now no longer create it. This saves about ~800 MB of memory.