Add Screen AI service for accessibility annotations.

    To improve the accessibility tree based on visual data, a snapshot of
    the screen is sent to a local machine learning library. The library will
    provide annotations for the snapshot and the annotations will be used
    to update the accessibility tree.
    This CL adds a Screen AI sand-boxed service to load the library, and
    general infrastructure to pass the images from renderer to the library
    and getting back and applying the annotations.
    The actual details on how the annotations are processed and used will be
    added in subsequent CLs.

    See more in go/chrome-screen-ai.

    This change is behind kScreenAI flag and disabled by default.