fix(memory): skip file content parts unsupported by the target model in history replay

[engmohamedsalah]

Fixes #41059

Summary

When a conversation's second turn arrives, TokenBufferMemory replays earlier turns and re-sends every past attachment as a native prompt content part — regardless of what the invoked model actually supports. On models whose provider does not implement document input (e.g. Qwen served via vLLM), the request fails with:

400 Unsupported chat content part type: 'file'

This change gates each converted history file part on the model schema's declared features using the existing AIModelEntity.supports_prompt_content_type() primitive:

• Supported content types → unchanged, sent natively
• Unsupported content types (e.g. documents replayed to a text/vision-only model) → degraded to a text note ([Unsupported file type: document]) so earlier turns keep contributing context instead of breaking the request
• Unresolvable model schema → previous send-anyway behavior preserved

Current-turn files are untouched; this only affects history replay.

Testing

• 3 new unit tests in tests/unit_tests/core/memory/test_token_buffer_memory.py:
  • unsupported document parts are replaced with text (regression for this issue)
  • supported image parts are still sent when VISION is declared
  • unresolved model schemas keep the old behavior
• Full file suite: 61 passed
• ruff check + ruff format --check clean