Wrap the live ICudaEngine after compile instead of serialize/deserialize. - #4698

Summary

• After build_engine_with_config, in-process convert_module wraps the live ICudaEngine instead of serialize() + C++ deserializeCudaEngine.
• interpret_module_to_result(..., serialize_engine=False) keeps cuda_engine on SerializedInterpreterResult. Cache insert, convert_exported_program_to_serialized_trt_engine, and get_extra_state / save still serialize from the live engine.
• Cross-compile-for-Windows still serializes at interpret time. In-process compile uses Python TRTEngine.execute() until save/reload (C++ execute path needs bytes).
• Klein NVFP4 cold compile: ICudaEngine.serialize 1.15s → 0, setup_engine 2.40s → 0.001s, compile wall 36.95s → 33.17s. First image ~1.12s and steady ~0.43s unchanged.

PR 10 of the compile-startup stack; based on #4693.

Type of change

• Performance (non-breaking for in-process compile; save/export still produce serialized engines)

Test plan

• tests/py/dynamo/runtime/test_live_engine_wrap.py (compile tiny Add: serialized_engine is None, live wrap, cosine match, get_extra_state serializes)
• Klein NVFP4 cold compile (torchtrt-213rc10): serialize 1.15s and setup_engine 2.40s gone; first image and steady unchanged
• Existing dynamo runtime tests (CI)

Checklist

• Code follows project style (pre-commit: isort, ruff, black, mypy)
• Self-review
• Comment on why the skip is valid
• Docs update (not needed)
• New tests
• Labels

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}