fix(ai21): update chat completions to use OpenAI-compatible Jamba API (/v1/chat/completions)

[cluster2600]

What

Updates the AI21 provider integration to route chat completion requests
to the new OpenAI-compatible /v1/chat/completions endpoint used by
the Jamba 1.5 / 1.6 model family, instead of the legacy
/v1/{model}/chat endpoint.

Also replaces the hand-rolled request/response format in
chatComplete.ts with the shared chatCompleteParams +
responseTransformers helpers from open-ai-base, exactly like the
Cerebras provider does.

Why

Fixes #1215.

The previous implementation targeted the legacy AI21 API that was
designed for older Jurassic-style models. AI21's current Jamba models
(jamba-instruct, jamba-1.5-mini, jamba-1.5-large, jamba-1.6-mini,
jamba-1.6-large) expose an OpenAI-compatible chat endpoint at
POST /v1/chat/completions. Sending requests to the old
/{model}/chat route returns a 404 for any Jamba model.

Reference: https://docs.ai21.com/reference/jamba-1-6-api-ref

How

src/providers/ai21/api.ts

• chatComplete case now returns /chat/completions (model is passed
  in the request body, as per the OpenAI spec).
• complete case is unchanged — legacy Jurassic-2 models still use
  /{model}/complete.

src/providers/ai21/chatComplete.ts

• Removed the old AI21ChatCompleteConfig (non-OpenAI param mapping),
  AI21ChatCompleteResponse interface, and
  AI21ChatCompleteResponseTransform.
• Kept only AI21ErrorResponseTransform (still used by embed.ts).

src/providers/ai21/index.ts

• chatComplete now uses chatCompleteParams() from open-ai-base,
  excluding the handful of OpenAI params the Jamba API does not support:
  logit_bias, logprobs, top_logprobs, service_tier,
  parallel_tool_calls.
• responseTransforms.chatComplete is provided by
  responseTransformers(AI21, { chatComplete: true }), which correctly
  stamps the provider name and delegates error handling to the standard
  OpenAI error transformer.

Testing

• TypeScript compiles without new errors (npx tsc --noEmit — pre-existing errors in bytez and open-ai-base are unchanged)
• npm run build succeeds (build log shows only pre-existing warnings)
• Manual curl against a local gateway with an AI21 API key against a Jamba model returns a valid chat completion response

Checklist

• Follows existing provider pattern (matches Cerebras, Groq, etc.)
• Backward-compatible: legacy Jurassic-2 complete and embed endpoints untouched
• No new TypeScript errors introduced
• Prettier formatting applied