MBS-10581: Build new prompt template system

[PM84]

Fixes #63

This PR introduces a structured prompt template system for the AI Text question type, replacing the legacy [[placeholder]] syntax with a more flexible {{placeholder}} system. The new architecture provides clear separation between admin-configured templates and question-specific content.

Changes

New Features

• Structured Prompt Template System: Admins can now configure a central prompt template with placeholders ({{role}}, {{questiontext}}, {{aiprompt}}, {{markscheme}}, {{defaultmark}}, {{response}}, {{jsonprompt}}, {{language}})
• Role Prompt Setting: New admin setting for defining the AI's role/persona
• Expert Mode: Question authors can take full control of the prompt by including {{response}} in their AI prompt, bypassing the admin template
• Expert Mode Template Button: UI button in the question edit form to easily insert the admin template for customization
• Confirmation Dialog: Detailed explanation of expert mode with all available placeholders when activating

Migration

• Automatic database migration (upgrade step 2026020601) converts legacy prompts:
  • [[expert]] → removed (was only a mode indicator)
  • [[response]] → {{response}}
  • [[questiontext]] → {{questiontext}}
  • [[userlang]] → {{language}}

Testing

• All 39 unit tests passing
• Codechecker clean
• Manual testing of:
  • Standard mode prompt generation
  • Expert mode with {{response}} detection
  • Expert mode template button and confirmation dialog
  • Legacy prompt migration

Breaking Changes

Questions using the old [[expert]]/[[response]] syntax will be automatically migrated during upgrade. No manual intervention required.

[marcusgreen]

Reading through code

[marcusgreen]

I have started reading through this and done a few tests, it looks good. I like the replacement of [[ ]] with {{ }}. The former was a legacy of me working with Moodle question types whereas the new approach is in line with the way most people deal with LLM systems. My tests so far have shown an unexpected improvement in responses, probably because the prompts are more like what an LLM expects. Back to testing.