[fork] hydra_config inline T3 + hydra_metrics response

[ddvnguyen]

Tracks the C++ side of the hydra_config / hydra_metrics implementation.

What

• Parse optional hydra_config from chat completion request body
• Compare model_path / split_mode / tensor_split against current engine state
• Trigger inline T3 reload if config differs (first-load or reload)
• Emit hydra_metrics in every chat completion response (model_path, split_mode, tensor_split, GPU memory, t3_reloaded, t3_reload_ms)
• Handle empty-start: engine starts without model, loads on first hydra_config
• split_mode / tensor_split added to server_context_meta

Files changed

• tools/server/server-context.cpp — hydra_config parsing, first-load path, T3 rebuild, metrics
• tools/server/server-context.h — meta fields, friend struct
• tools/server/server-task.cpp — emit hydra_metrics in response
• tools/server/server-task.h — hydra_metrics field
• tools/server/server.cpp — empty model path handling
• src/llama-hydra.cpp — split_mode/override_tensor staging without ctx

Hydra parent issue

ddvnguyen/hydra_vortex#TBD