Implement W&B Inference adapter

[op-simoneromeo]

Summary

• replace the ai-wandb stub with a W&B Inference OpenAI-compatible chat-completions adapter
• use the documented https://api.inference.wandb.ai/v1/chat/completions route with WANDB_API_KEY
• support dry-run, system/user messages, max_tokens/temperature, stream=false, extra passthrough params, usage mapping, text-choice fallback, custom baseUrl, optional OpenAI-Project usage tracking, and provider error excerpts
• add mocked fetch tests for dry-run, request shape, response mapping, compatible text choices, and error handling

Validation

• corepack pnpm@9.12.0 exec vitest run packages/ai/wandb/src/index.test.ts
• corepack pnpm@9.12.0 --filter @profullstack/sh1pt-ai-wandb typecheck
• corepack pnpm@9.12.0 --filter @profullstack/sh1pt-ai-wandb build
• git diff --check
• common secret-pattern scan over changed W&B files

No real W&B API key or other production secret was used; tests mock fetch.