fix(ai): accept content part arrays in OpenAI Chat streams

[kitlangton]

Why

Mistral-style models served through OpenAI-compatible /chat/completions gateways stream choices[0].delta.content as an array of typed content parts instead of a plain string:

{
  "id": "chunk_fixture",
  "object": "chat.completion.chunk",
  "created": 1791353545,
  "model": "example-model",
  "choices": [
    {
      "index": 0,
      "finish_reason": null,
      "logprobs": null,
      "delta": {
        "content": [
          {
            "type": "thinking",
            "thinking": [{ "type": "text", "text": "Let" }]
          }
        ]
      }
    }
  ]
}

Because OpenAIChatDelta.content only accepted string | null | undefined, every run against these models failed on the first chunk with InvalidProviderOutput.

What Changes

OpenAIChatDelta.content now accepts string | null | ReadonlyArray<OpenAIChatContentPart>, and contentDeltas flattens part arrays in order:

delta.content input	Stream behavior
"Hello"	Emits text-delta ("Hello") on text-0
[{ "type": "text", "text": "Hello" }]	Emits text-delta ("Hello") on text-0
[{ "type": "thinking", "thinking": [{ "type": "text", "text": "Let" }] }]	Emits reasoning-delta ("Let") on reasoning-0
[{ "type": "thinking", "thinking": "Planned." }]	Emits reasoning-delta ("Planned.") on reasoning-0
Other part types (reference, image_url, …)	Skipped and logged at debug level (openai-chat.content_parts_skipped)
[] after finish_reason	Accepted as an empty trailing chunk
Non-empty text or thinking parts after finish_reason	Rejected with InvalidProviderOutput (OpenAI Chat received content after the finish reason)

Verification

bun test packages/ai/test/provider/openai-compatible-chat.test.ts
bun run --cwd packages/ai test
bun run --cwd packages/ai typecheck
bunx oxlint packages/ai/src/protocols/openai-chat.ts packages/ai/test/provider/openai-compatible-chat.test.ts

• openai-compatible-chat.test.ts: 25 pass, 0 fail (including 5 new tests under content part arrays covering thinking/text streaming and replay, intra-chunk ordering, skipped/empty parts, and post-finish validation).
• @opencode/ai test suite: 1,782 pass, 0 fail.
• @opencode/ai typecheck (tsgo --noEmit && tsgo --noEmit -p tsconfig.types.json) and workspace turbo typecheck: pass.
• oxlint on changed files: 0 warnings, 0 errors.