TICKET-611: Add openai-remote provider for the ai-telemetry-gateway

[Naragod]

Cloud OpenAI calls from MarkUs grading must pass through the ai-telemetry-gateway so every call is attributed to a course and checked against budgets. No existing provider can reach it: openai talks straight to api.openai.com, and remote speaks polymouth's custom contract. This adds the provider for the gateway path; MarkUsProject/markus-autotesting#731 is the autotester side.

• Register openai-remote ->OpenAIRemoteModel, so the autotester selects the gateway path like any other model.
• Subclass OpenAIModel, changing only the endpoint (--remote_url, the gateway's OpenAI-compatible /v1 root) and auth (LITELLM_API_KEY virtual key as Authorization: Bearer), since the gateway speaks the standard OpenAI contract.
• Forward the x-litellm-spend-logs-metadata header from LITELLM_SPEND_METADATA (set per invocation by the autotester), failing loud on malformed JSON, because the gateway rejects unattributed calls.
• Default max_tokens to 1024 (caller values win), because the gateway rejects calls that omit it and the tester config has no field to set one.

Known limitation: scope: image uses process_image, which does not send max_tokens, so the gateway rejects it; code and text feedback are the validated paths.

[donny-wong]

Thanks @Naragod , looks good.