feat: add text_extraction activity type

[siculo]

Summary

Add text_extraction as a new activity type. Where tag_extraction
generates semantic tags describing the content of a photo,
text_extraction captures literal textual content visible in the
image — signage, billboards, menus, labels, street signs, documents,
text on buildings.

Motivation

Vision models already in use (qwen2-vl and similar) are capable of
reading text from images without additional providers or specialized
OCR models. The use case emerged from observing that a photo of a
billboard produces semantic tags ("billboard", "advertising") but
the actual text on the sign is never captured.

Text extraction adds a qualitatively different metadata layer that
complements tag extraction and can be run independently or in
sequence on the same project.

Scope

API

• Add text_extraction as a valid value for the activity_type
  field in POST /api/projects/{id}/activities
• Define the prompt used for text extraction (instruct the model to
  return only textual content found in the image, not a description)
• Define the output format for extracted text (list of strings,
  or single block, to be decided)
• Return 400 with unsupported_activity_type for unknown types
  (already specified in issue feat: update API paths and request shapes to new terminology, update plugin accordingly #92 — no new error code needed)

Plugin

• Support text_extraction in the activity creation modal:
  • Add a "New Text Extraction" entry point (button or menu item,
    consistent with the pattern defined for future activity types
    in issue feat: update Lightroom plugin UI to reflect new terminology #93)
  • Apply extracted text to Lightroom metadata fields
    (keywords, caption, or description — to be decided)

Documentation

• Update API reference with the new activity type and output format
• Update evolution document (already done in docs/)

Considerations

• Activity context (issue feat: add context field to activity creation #122) is particularly useful here: the
  user can guide the model ("look for billboard text",
  "extract menu items") to improve extraction quality
• Mixed-language text is handled naturally by multilingual models
• Output quality depends on image resolution and font legibility

Acceptance criteria

• POST /api/projects/{id}/activities with
  activity_type: "text_extraction" creates and queues
  an activity successfully
• Extracted text is returned in the activity results
• Plugin applies extracted text to the appropriate Lightroom
  metadata field
• Existing tag_extraction activities are unaffected