feat(providers): add Cheaper Inference as an OpenAI-compatible gateway

[aiapienthusiast]

Description

This PR adds Cheaper Inference as a provider template.
Cheaper Inference is an OpenAI-compatible LLM gateway.
It gives access to models from several labs with one API key.

Cheaper Inference is one of the fastest-growing AI routers.
Each model costs 15–60% less than the list price of its lab.

The template uses the generic openai_compat backend.
It follows the pattern of #159.

Related Issues

None.

Changes Made

• Add a cheaperinference ProviderSpec with the gateway endpoint class.
• Add the matching ProvidersConfig field.
• Add a setup section to docs/guide/models.md.
• Add unit tests.

Checklist

• Changes tested locally
• Code reviewed
• Documentation updated (if necessary)
• Unit tests added (if applicable)

Additional Notes

Model IDs are bare, so the spec sets strip_model_prefix=True.
The tests make no network calls.
I checked the API with a live key: model list, chat, streaming, tool calls and JSON schema output with gpt-5.4-mini.

Details

Spec values:

• Base URL: https://api.cheaperinference.com/v1
• Key variable: CHEAPER_INFERENCE_API_KEY
• Base detection keyword: cheaperinference.com
• Model catalog: auto, which resolves to the OpenAI /v1/models listing.

Files:

• core/providers/registry.py
• core/config.py
• docs/guide/models.md
• tests/test_cheaperinference_provider.py (new)
• tests/test_provider_egress.py

Commands run:

python -m pytest -q tests/test_cheaperinference_provider.py tests/test_provider_egress.py <related provider tests>
# 68 passed
python -m pytest -q <all provider, catalog, config, profile and connection test files>
# 217 passed
ruff check --ignore=E402 core/config.py core/providers/registry.py tests/test_cheaperinference_provider.py tests/test_provider_egress.py
ruff format --check core/config.py core/providers/registry.py tests/test_cheaperinference_provider.py tests/test_provider_egress.py
# ruff 0.15.21, as in .pre-commit-config.yaml; all passed

Cheaper Inference has no embeddings endpoint. The docs say so.