Hybrid Engine OPT cache compatibility

Hybrid Engine’s injected OPT layers implement the legacy tuple-cache interface. OPT decoder layers exposing cache_position or past_key_values use a newer cache contract. Those models now retain native Hugging Face generation, with an explicit warning that inference acceleration is unavailable. Training and engine.module.generate() remain available; this does not add native-kernel support for the newer Cache interface.

The legacy OPT injection path is unchanged. The fallback does not provide inference tensor parallelism, continuous-batching native-cache operations, or CUDA Graph acceleration. Use the supported legacy path when those features are required.

The offline regression in tests/unit/hybrid_engine/test_he_opt_cache.py uses two data-parallel ranks, a tiny randomly initialized OPT, and repeated training/generation transitions. Each rank compares greedy output tokens with an independent Hugging Face model loaded from the updated weights. No model download is required.