Model overlays: base model + merged overrides (generalize per-pipeline overrides)

[localai-bot]

Motivation

Realtime pipelines have started growing per-stage override fields to tweak the contained models without editing their configs:

• pipeline.reasoning_effort (feat: forward reasoning_effort to the backend so jinja models honor it #10184, merged)
• pipeline.disable_thinking (feat(realtime): stream the LLM / TTS / transcription pipeline stages #10176)

As richiejp noted in #10176, this "creates two layers of config" and doesn't scale — every new knob needs a new pipeline field. A generic mechanism would be cleaner and useful well beyond realtime.

Proposal: model overlays

Introduce a model overlay: a model config that inherits from a base model and merges its own overrides on top.

name: qwen3-no-think
base: qwen3-4b          # inherit everything from qwen3-4b
reasoning:
  disable: true         # override just this

A realtime pipeline (or any caller) then just points at the overlay:

name: gpt-realtime
pipeline:
  llm: qwen3-no-think

This subsumes pipeline.reasoning_effort / pipeline.disable_thinking (and any future per-pipeline overrides) with one mechanism: one base model, N overlays inheriting base settings and merging overrides. Users get quick per-model "profiles" without duplicating configs.

Notes / open questions

• Merge semantics: scalar override wins; how to handle maps/slices (replace vs merge)?
• Resolve overlays at config-load time so the rest of the stack sees a fully-merged ModelConfig.
• Cycle detection on base.
• Once overlays exist, the targeted pipeline.* override fields could be deprecated in favor of overlays.

Follow-up from PR #10176 / #10184.