fix(tools): await conversation persistence

[GautamSharma99]

Summary

The Vercel AI SDK middleware previously started conversation persistence without
awaiting it. This allowed non-streaming requests and output streams to finish
while the /v4/conversations request was still pending.

In serverless environments, the runtime could terminate immediately afterward and
silently lose the conversation.

Changes

• Await conversation persistence before doGenerate resolves.
• Await persistence during stream finalization.
• Preserve immediate delivery of streamed response chunks.
• Add regression tests for both non-streaming and streaming paths.

Testing

• Focused Vitest tests pass.
• Focused Bun tests pass.
• Biome checks pass.
• Package build passes.
• Changed files have no TypeScript diagnostics.
• Full repository suite: 311 passed, 64 skipped, with 21 pre-existing unrelated
  failures.

Notes

Persistence errors remain non-fatal because saveMemoryAfterResponse already
handles and logs failures internally.