feat: add Infron AI chat provider

[meridah7]

What

Adds Infron as a provider for the puter-chat-completion driver, following doc/contributing-apis.md and using the OpenRouter provider as the reference implementation (as suggested by the Puter team).

Infron is an AI inference platform with an OpenAI-compatible gateway (https://llm.onerouter.pro/v1) serving 400+ models across 100+ upstream inference providers.

How it works

• InfronProvider mirrors OpenRouterProvider: OpenAI SDK pointed at the gateway, dynamic model catalog from GET /v1/models (kv-cached for 15 minutes). Two deliberate differences, both driven by the gateway's actual API shape:
  • The catalog endpoint requires authentication, so the fetch sends the configured key as a bearer token.
  • The catalog has no OpenRouter-style pricing object; prices are USD per million tokens (min_prompt_price / min_completion_price), converted to microcents per token (× 100). Non-chat catalog entries (image/video/embeddings/search) and display-only entries are filtered out via category_type / is_display_only.
• Metering: requests always send usage: { include: true }, which makes the gateway report an authoritative cost (USD) on the response. Like OpenRouter's usage.cost branch, that cost is billed as a single billedUsage line item with per-token costs zeroed. One wrinkle: Infron reports cost at the top level of the response rather than inside usage — non-streaming reads it off the spread completion, streaming merges it in via an index_usage_from_stream_chunk deviation. If cost is ever absent, metering falls back to catalog per-token pricing (cache reads billed at the full prompt rate there, so the fallback never under-bills).
• Registration: config key infron (apiKey + optional apiBaseUrl), added to the AGGREGATORS set so its aliases never shadow first-party providers in the model map — same treatment as openrouter and together-ai.
• Models are exposed as infron:<vendor>/<model> with the same alias scheme as OpenRouter. Default model: infron:deepseek/deepseek-v4-flash.

Testing

• Unit tests (InfronProvider.test.ts, 13 tests): same harness as the OpenRouter suite — real PuterServer + live MeteringService, OpenAI SDK and axios mocked at the module boundary. Covers construction/base-URL override, authenticated catalog fetch, kv caching, category filtering, price-unit conversion, request shape (prefix stripping, usage.include, stream_options), both metering branches (top-level cost and per-token fallback, non-stream and stream), and moderation.
• Integration test (InfronProvider.integration.test.ts): env-gated on PUTER_TEST_AI_INFRON_API_KEY, exercises the real gateway; verified passing locally.
• Full src/backend/drivers/ai-chat/ suite passes locally (428 tests).
• Response shapes (models catalog schema, top-level cost on stream + non-stream, usage field names) were all verified against the live gateway while writing this, not assumed from docs.

Security pass

• The API key leaves the process only toward the configured gateway (SDK auth header + catalog bearer token); it is never logged.
• No new HTTP surface; provider is inert unless providers.infron.apiKey is configured.
• Errors from the gateway propagate through the existing driver classification path unchanged.

Happy to adjust anything to better match house conventions!

[CLAassistant]

All committers have signed the CLA.

[ProgrammerIn-wonderland]

.

[Salazareo]

@meridah7 could you sign the CLA please