Support disable metal buffer cache to prevent performance degradation caused by large memory caching (#390)

* support disable metal buffer cache, due to large unused memory buffered when llm generated long context tokens

* Run format and add "cache_enabled" feature tests

python/src/metal.cpp

@@ -11,4 +11,12 @@ using namespace mlx::core;
 void init_metal(py::module_& m) {
   py::module_ metal = m.def_submodule("metal", "mlx.metal");
   metal.def("is_available", &metal::is_available);
+  metal.def(
+      "cache_enabled",
+      &metal::cache_enabled,
+      "check if metal buffer cache is enabled, default is true");
+  metal.def(
+      "set_cache_enabled",
+      &metal::set_cache_enabled,
+      "enable or disable metal buffer cache");
 }