[Feature]: support multimodal message ingestion with transcript-first rawdata storage #4

Is your feature request related to a problem? Please describe.
Memind currently assumes text-only messages. The current framework-neutral message model (ChatMessage) only carries role + String content, which means image and audio messages cannot enter the memory pipeline directly. In real chat applications, a large part of useful memory comes from screenshots, photos, voice notes, and other multimodal inputs. Today, users would need to pre-process those assets outside memind and manually convert them to text before ingestion.

Describe the solution you'd like
Add first-class multimodal message ingestion so memind can accept image and audio inputs, convert them into normalized text, and persist both the transcript and the original asset references in a way that fits the existing extraction pipeline.

Suggested scope for this issue:

• extend the ingestion/message model to represent text, image, and audio content instead of only plain text
• add a multimodal ingestion pipeline that can run OCR / captioning for images and ASR / transcription for audio
• persist the normalized text output into rawdata so the existing rawdata -> item -> insight pipeline can continue to work without requiring a parallel retrieval path
• store references to the original multimodal assets alongside the derived text, metadata, and processing provenance
• introduce a pluggable asset storage abstraction so original image/audio resources can be stored either on local filesystem storage or in S3-compatible object storage
• expose starter / configuration support so applications can choose storage backend and multimodal parsing providers

Recommended functional behavior:

• image input: store original asset, extract OCR/caption text, then write normalized text into rawdata
• audio input: store original asset, run transcription, then write normalized text into rawdata
• rawdata should remain the canonical entry point for downstream extraction, with links back to the original asset location and media metadata

Out of scope for the first iteration:

• full video ingestion
• native binary retrieval without transcript
• provider-specific UI concerns

Describe alternatives you've considered

• Require upstream applications to convert image/audio to text before calling memind. This is simpler, but it pushes a core memory-ingestion responsibility to every integrator and makes multimodal behavior inconsistent across applications.
• Only store image/audio metadata and asset references without transcript extraction. This preserves the files, but the current memory extraction and retrieval pipeline still cannot reason over the multimodal content.
• Store binary payloads directly inside rawdata / database rows. This would complicate storage backends and is less portable than using asset references plus transcript-first indexing.

Additional context
This issue can act as an umbrella for follow-up sub-issues such as:

• multimodal message model / API changes
• asset storage SPI with local and S3 implementations
• OCR / ASR provider abstraction
• rawdata schema changes for media references and provenance
• Spring Boot starter auto-configuration and examples