Tighten image_generation defaults and refuse streaming requests

Code review surfaced four real problems with the new image_gen path:

* Default `enable_image_gen` was True, which silently injected the
  image_generation tool into every chat completion. That let the model
  choose to draw a picture for prompts that never asked for one — 25s
  extra latency and subscription quota burn for what should have been
  plain text. Default is now False; opt in with CODEX_ENABLE_IMAGE_GEN=1.
* The streaming chat_completions consumer only handles response.output_text
  events. With image_gen enabled and stream=true, the model emits
  image_generation_call events the streaming loop drops on the floor —
  the client gets an empty stream. Refuse the combination at the entry
  point with HTTP 400 + a clear error message.
* `revised_prompt` from the model was used as markdown alt text without
  escaping. A `]` or `)` in there would break the `![alt](url)` syntax
  and let the base64 payload leak out as visible text. Escape `[]()\`
  and clamp alt to 200 chars.
* `_truncate_for_log` now replaces `data:image/...;base64,...` URIs with
  a short `<image/png NNN B base64 omitted>` placeholder before length
  truncation. Without that, every full/qa-mode response log starts with
  3.9KB of useless base64 prefix and the actual text gets clipped.