llama : fix command-r detokenization

Author: compilade

src/llama.cpp

@@ -5407,6 +5407,7 @@ static void llm_load_vocab(
             } else if (
                 tokenizer_pre == "command-r") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_COMMAND_R;
+                vocab.tokenizer_clean_spaces = false;
             } else if (
                 tokenizer_pre == "qwen2") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_QWEN2;