feat: azure adapter and OpenAI Files API uploads

[Im-Beast]

• Adds azureOpenAIModel, exported as @alphaxiv/agents/azure-openai, sharing the Open Responses adapter and reasoning resolution with openAIModel
• Adds a fileStore option to both: images and PDFs are uploaded to the provider Files API once and sent as file_id afterwards instead of base64 on every call
• The store is caller owned and maps a URL to a file id for 7 days; set returns the winning id so racing uploads agree, delete only drops a row still holding the id, and expiry sweeping is the caller's job
• A rejected file id is handled inside the adapter stream: a stored id is dropped and re-uploaded, an id uploaded in the same call is resent unchanged after a pause, three sends at most, so the Agent retry loop never sees it
• Adds the request_start stream item, which restarts the first token watchdog after uploads
• Adds ActiveTrace.update for partial trace content
• Live tests against OpenAI and Azure are skipped without provider keys

Measured on gpt-5.6-luna with a 2 MB PDF, 25 prompts per arm on my mac, so the download and upload might be slower than on a real deployment.

TTFT:

provider	arm	median	p90	mean
OpenAI	base64	3875	4799	3939
OpenAI	store, warm	2849	3481	2871
OpenAI	store, cold	4601	5612	4732
Azure	base64	3929	5763	4297
Azure	store, warm	2717	3493	2929
Azure	store, cold	4722	8240	5432

Total:

provider	arm	median	p90	mean
OpenAI	base64	4629	6066	4784
OpenAI	store, warm	3555	4361	3654
OpenAI	store, cold	5342	6803	5486
Azure	base64	4385	6982	4765
Azure	store, warm	3255	4431	3483
Azure	store, cold	5132	8607	6090