Kv cache (#643)

* in place kv_cache

* fix

* fix kv cache size

* partially fix kv cache dtype

* step kv cache

* multiple of step size

* more teests + kv cache

* more kv cache

* udpate all models to use kv cache

llms/mlx_lm/tokenizer_utils.py

@@ -314,7 +314,8 @@ def load_tokenizer(model_path, tokenizer_config_extra={}):
 
     tokenizer_file = model_path / "tokenizer.json"
     if tokenizer_file.exists():
-        tokenizer_content = json.load(tokenizer_file.open())
+        with open(tokenizer_file, "r") as fid:
+            tokenizer_content = json.load(fid)
         if "decoder" in tokenizer_content:
             if _is_spm_decoder(tokenizer_content["decoder"]):
                 detokenizer_class = SPMStreamingDetokenizer