test: refactor tests and fix inference edge cases

- Convert protocol and MoE test classes to plain functions
- Add real server/engine integration and generate_async tests
- Isolate test model per test and use pytest tmp_path
- Reset FastAPI engine state after inference tests
- Fix generate_async StopIteration handling on Python 3.12
- Fix HF adapter MoE dense/shared and Gemma qk_norm mapping
- Correct dev dependency httpx2 to httpx

tests/conftest.py

@@ -1,7 +1,5 @@
 import json
 import os
-import shutil
-import tempfile
 
 import pytest
 import torch
@@ -44,20 +42,18 @@ def test_tokenizer():
     return create_test_tokenizer()
 
 
-@pytest.fixture(scope="session")
+@pytest.fixture
 def test_model(device):
-    """Session-scoped small AutoRegressiveLM model, created once."""
+    """Function-scoped small AutoRegressiveLM model, isolated per test."""
     config = make_tiny_config()
     model = AutoRegressiveLM(config).to(device=device)
     return {"model": model, "device": device, "config": config}
 
 
 @pytest.fixture
-def temp_dir():
-    """Function-scoped temporary directory, cleaned up after each test."""
-    d = tempfile.mkdtemp()
-    yield d
-    shutil.rmtree(d, ignore_errors=True)
+def temp_dir(tmp_path):
+    """Function-scoped temporary directory, cleaned up by pytest."""
+    return str(tmp_path)
 
 
 @pytest.fixture