fix: torch-TRT runtime cache attribute + standard-TRT fast refit regression

[tp5uiuc]

Summary

Two related fixes to issues introduced by recent torch-TensorRT runtime-cache and refit work, both surfaced by the L2 dynamo compile tests.

• fix(runtime): preserve runtime_cache/runtime_config on rehydration paths — PythonTorchTensorRTModule.__getstate__ pops runtime_config and runtime_cache (they hold non-picklable native handles), but __setstate__ never restored them. setup_engine() only re-creates them inside _setup_runtime_config(), which is gated by ENABLED_FEATURES.tensorrt_rtx. On standard TRT the gate is false, so on every unpickle / deepcopy the attributes simply do not exist — __del__ -> _save_runtime_cache() then reads self.runtime_cache and Python emits PytestUnraisableExceptionWarning across the refit and weight-stripped-engine test suites. Re-init both fields to None in __setstate__ and _load_from_state_dict before setup_engine(). No behavior change on TRT-RTX.
• fix(refit): scope unset-weights strict check to TRT-RTX only — the strict cross-check unset_weights = {w for w in weight_list if w not in mapping} added previously broke fast refit on standard TRT for any engine with CONSTANT layers intentionally absent from the mapping (e.g. batch-norm eps constants baked at build time). The pre-existing warn-and-continue branch makes that absence the contract; the strict check made it an error. On a resnet18 disk-engine-cache hit, the path now asserts with "0 missing, 20 unset" and _pretraced_backend silently falls back to GraphModule (visible as XPASS on test_dynamo_compile_with_default_disk_engine_cache and test_torch_compile_with_default_disk_engine_cache). The strict check exists only to guard the TRT-RTX case where each weight lives in its own independent wtsEngine and get_missing_weights() can under-report; on standard TRT, get_missing_weights() is authoritative because connected weight engines surface any unset weight transitively. Gate the assertion behind ENABLED_FEATURES.tensorrt_rtx.
• chore(refit): drop unused CPU_DEVICE import — pre-existing F401 surfaced by ruff on the touched file.

Test plan

Verified end-to-end on a fresh standard-TRT (non-RTX) install of nightly torch_tensorrt on Linux + CUDA 13.0:

• Fix 1: pytest tests/py/dynamo/models/test_model_refit.py::test_refit_one_engine_with_weightmap test_refit_one_engine_python_runtime_with_weightmap test_complex_buffer_with_real_param_refit — baseline reproduces 2× AttributeError: ...runtime_cache; with patch, 0 occurrences.
• Fix 2: pytest tests/py/dynamo/models/test_engine_cache.py::TestEngineCache::test_dynamo_compile_with_default_disk_engine_cache test_torch_compile_with_default_disk_engine_cache — baseline reproduces 1× AssertionError: Fast refitting failed due to incomplete mapping (0 missing, 20 unset) plus Returning GraphModule forward instead silent fallback; with patch, 0 of either, TRT compilation actually succeeds on the cache-hit path.
• CI sweep on Windows L2 dynamo compile job to confirm the original PytestUnraisableExceptionWarning groups disappear.

🤖 Generated with Claude Code

[zewenli98]

LGTM