refactor: extract MTMDBaseHandler to prepare for TTS support

- Centralize shared configuration, context initialization, capability
  detection, media loading, and resource cleanup in MTMDBaseHandler.
- Keep chat templates, message parsing, and text generation in
  MTMDChatHandler while preserving existing constructor compatibility.
- Reuse common infrastructure for a future MTMDAudioGenerator, avoiding
  duplicated initialization and media handling across chat and TTS.
- Release dependent resources before the mtmd context and ensure context
  cleanup still runs when a resource callback fails.
- Add regression coverage for shared initialization, template compatibility,
  subclass media loading, and cleanup behavior.

Signed-off-by: JamePeng <jame_peng@sina.com>

Author: JamePeng