feat: Add multi-modality support for content blocks in PrecisePrefixCacheScorer

[guygir]

IMPORTANT: Depends on llm-d/llm-d-kv-cache#255. This PR should not be merged until the kv-cache PR is merged first.

This PR extends the PrecisePrefixCacheScorer to preserve structured multi-modality content blocks (OpenAI API format) through the pipeline.
This is the first stage of multi-modality support - only basic technical feasibility. This PR focuses solely on images - not audio or video, because GAIE already supports images but audio/video support requires additional GAIE changes (These will be addressed in the next stage)

Changes:

• Added convertContentToPreprocessingFormat() helper to convert GAIE's structured content blocks to OpenAI API format
• Modified getScores() to preserve structured content instead of using raw text
• Maintains backward compatibility with text-only content

What Works:

• Multi-modal requests are correctly parsed and preserved from GAIE
• Structured content blocks flow correctly through the scheduler
• Chat template rendering works with structured content
• Requests are forwarded correctly to vLLM
• Backward compatible with text-only requests

Known Limitations (for current stage):

• Tokenization may not match vLLM exactly (images tokenized as text, not vision tokens - this only affects merged preprocessor models like Qwen2-VL)
• Block hashes may not match vLLM exactly (missing mm_hash - we're consistent with ourselves but won't match vLLM's hashes for multimodal blocks, which is not an issue, just FYI)
• These will be addressed in the next stage.

Testing:

• Multi-modality conversion logic test (test/multi_modality/test_multimodality_conversion.go)