feat: unified response_format / structured JSON output support

[stackbilt-admin]

Summary

Add a unified response_format option to LLMRequest that enforces structured JSON output across all providers.

Motivation

From the frontend team (Oracle project): they need structured JSON output (a files[] array) and currently have no way to enforce this through the package. Each upstream provider handles this differently:

• OpenAI: response_format: { type: "json_object" }
• Anthropic: prefilled assistant responses / tool use
• Cloudflare Workers AI: no native support
• Groq: response_format: { type: "json_object" }
• Cerebras: varies by model

Proposed API

const response = await llm.chat({
  provider: 'openai',
  model: 'gpt-4o-mini',
  messages: [{ role: 'user', content: 'List the files as JSON' }],
  response_format: { type: 'json_object' },
});

Implementation approach

Each provider adapter translates the unified response_format to its native mechanism:

Provider	Native mechanism
OpenAI	response_format: { type: "json_object" } (pass-through)
Groq	response_format: { type: "json_object" } (pass-through)
Anthropic	Prefilled assistant turn { + system prompt instruction
Cerebras	System prompt instruction + validation
Cloudflare	System prompt instruction + validation

For providers without native support, inject a system prompt suffix and optionally validate the response parses as JSON.

Acceptance criteria

• response_format field added to LLMRequest type
• OpenAI and Groq pass through natively
• Anthropic uses prefilled assistant response approach
• Cloudflare and Cerebras use system prompt injection fallback
• Optional JSON parse validation on response
• Unit tests for each provider's translation
• Documented in README

[stackbilt-admin]

Shipped in v1.0.0 (commit 743a1e0). Unified response_format: { type: 'json_object' } supported across OpenAI, Anthropic, Cloudflare, Cerebras, and Groq.