Allow disabling chat completions persistence via null store

[meffmadd]

What does this PR do?

Allow operators to disable chat completion persistence by setting storage.stores.inference to null in a run config. When the reference is None, the auto-router factory in get_auto_router_impl skips constructing and initializing the InferenceStore entirely: no inference_store table is created and no background write workers are started, so no chat completion payload is ever persisted.

Behavior when persistence is disabled:

• Chat completions continue to work for both streaming and non-streaming requests.
• The history endpoints (list, retrieve, messages) raise NotImplementedError, which the exception mapping translates to HTTP 501 — rather than returning an empty list or a 404.
• Other stores (responses, datasets, eval, files, prompts, vector_io) stay enabled, so disabling inference persistence is independent of the rest of the storage layer.
• The default remains backward compatible: omitting the inference key keeps persistence enabled; only an explicit inference: null disables it.

This follows the same optional-store pattern the Responses store already uses.

Configuration

storage:
  stores:
    inference: null

Changes

• src/ogx/core/routers/__init__.py — build/initialize the InferenceStore only when the reference is not None.
• src/ogx/core/storage/README.md, src/ogx/distributions/README.md — document the opt-out and its runtime behavior.
• Unit tests:
  • tests/unit/core/routers/test_inference_router.py — non-streaming and streaming completions with no store; list/get/messages raise NotImplementedError.
  • tests/unit/core/test_storage_references.py — omit-vs-null parsing: omitted key keeps the enabled default, null disables.
  • tests/unit/server/test_resolver.py — resolver does not construct/initialize an InferenceStore when the reference is None.
• Integration test tests/integration/inference/test_inference_store_disabled.py (recorded responses from openai/gpt-4o) — end-to-end behavior over the full HTTP path, including asserting that no inference_store table exists in the SQL backend.

Test Plan

Unit tests

uv run pytest tests/unit/core/routers/test_inference_router.py \
  tests/unit/core/test_storage_references.py \
  tests/unit/server/test_resolver.py -q

Output:

.........................                                [100%]
============================= slowest 10 durations =============================
0.08s call     tests/unit/server/test_resolver.py::test_resolve_impls_basic

(9 durations < 0.005s hidden.  Use -vv to show these durations.)
25 passed in 0.18s

Integration tests (replay mode — no API keys required)

uv run --no-sync ./scripts/integration-tests.sh \
  --stack-config server:ci-tests --setup gpt \
  --file tests/integration/inference/test_inference_store_disabled.py

Output:

tests/integration/inference/test_inference_store_disabled.py::test_non_streaming_chat_completion_without_store PASSED [ 16%]
tests/integration/inference/test_inference_store_disabled.py::test_streaming_chat_completion_without_store PASSED [ 33%]
tests/integration/inference/test_inference_store_disabled.py::test_list_chat_completions_reports_not_configured PASSED [ 50%]
tests/integration/inference/test_inference_store_disabled.py::test_retrieve_chat_completion_reports_not_configured PASSED [ 66%]
tests/integration/inference/test_inference_store_disabled.py::test_list_chat_completion_messages_reports_not_configured PASSED [ 83%]
tests/integration/inference/test_inference_store_disabled.py::test_no_inference_store_table_when_persistence_disabled PASSED [100%]

======================== 6 passed, 2 warnings in 6.69s =========================

Both suites pass on this branch. The integration tests run in replay mode against the committed recordings, so they are deterministic and need no API keys.