Decouple HF and vLLM framework logic from core monitoring engine #33

Part of #32

Problem

HF and vLLM specific logic is embedded in core engine files. Capacity check, step bytes computation, and step prep sequence are duplicated between generate.py (HF) and vllm_integration.py (vLLM). Adding a new framework requires modifying core code.

Current coupling

• generate.py (1168 lines) — entirely HF-coupled (StaticCache, CompileConfig, past_key_values, attention mask decoding)
• vllm_integration.py (434 lines) — entirely vLLM-coupled (scheduler_output parsing, CUDA graph padding, architecture remap, force_eager monkey-patch)
• engine.py _prepare_ring_step — hardcoded to HF kwargs (input_ids, attention_mask, past_key_values)
• ring_transport.py — vLLM-specific preset (vllm-full) and batch=0 flattened-batch handling mixed in

What to do

• Extract shared core logic (capacity check, step bytes calc, step prep sequence) into a unified interface in monitoring/engine.py
• Move HF-specific wrapper to integration/hf_adapter.py
• Move vLLM-specific wrapper to integration/vllm_adapter.py
• Adapters translate framework-specific inputs and call the unified core interface