Batches: report reasoning_tokens and the finalizing status

[JGoutin]

Two cosmetic divergences from the upstream batch API, grouped because they are the same
size:

• usage.output_tokens_details.reasoning_tokens is never populated, although the
  per-request usage carries it for reasoning models.
• The finalizing status is never reported, so a client charting batch progress sees
  in_progress jump straight to completed.

Neither breaks a client. Both are visible to one reading the objects side by side with
upstream's, which is what a compatibility claim invites.