feat(openai): record input and output images for image APIs

[satyadevai]

Summary

Records span-level images for OpenAI image generation:

• output.images.N.image.url for Images API results and Responses API image generation calls
• input.images.N.image.url for the files sent to image edits and variations

Before this, generated images were only in the raw output.value JSON, and Responses API image generation calls were ignored.

Changes

Images API images.generate, sync and async)

• Each item in data becomes output.images.N.image.url. A hosted url is used as is. b64_json is converted to a data URL.
• Image type for b64_json comes from the response output_format, then the request output_format, and defaults toimage/png.
• Edit and variation files in the image, image[] and mask fields become input.images.N.image.url. Supported inputs are bytes, file objects, paths and (ftype) tuples.
• Input image type comes from the tuple content type, then the file's magic bytes, then the filename, and defaults to image/png.
• File objects are read and seeked back to where they were. Non-seekable streams are skipped so the request body isn't consumed.
• For streaming generation and edits, the new _ImagesAccumulator records the image from the
  image_generation.completed or `image_edit.c

Responses API

• Each image_generation_call with a result becomes output.images.N.image.url, numbered in output order.
• Image type comes from the item's output_format of the image_generation tool in therequest.

Dependencies

• Raised the floor to openinference-instrumeease applies trace config masking(hide_inputs, hide_outputs, hide_input_images, base64_image_max_length) to span-level input.imagesandoutput.images`.