fix(backend): drop the prompt cache before the allocator-empty; note clear_cache is thread-agnostic

Review follow-ups: clear_voice_prompt_memory_cache() now runs before
backend.unload_model() at all three call sites so the device-resident
prompt tensors are already released when empty_device_cache() /
empty_mlx_cache() runs, instead of going back into the caching allocator
afterwards. empty_mlx_cache's docstring states why it may run off the
MLX worker thread.

backend/backends/base.py

@@ -232,6 +232,12 @@ def empty_mlx_cache() -> None:
     returning them to the OS. Backends must call this after unloading an
     MLX model, or the process's memory footprint never shrinks even though
     the model object itself was dropped.
+
+    Safe from any thread: ``mx.clear_cache`` only drains the global
+    allocator pool and never touches the per-thread stream registry, so
+    unlike load/generate it does not have to run on the MLX worker thread
+    (verified from the FastAPI event loop with a generation in flight on
+    the worker).
     """
     import mlx.core as mx