feat(0406): Add Gemma4 Unified vision-language support

[vbaddi]

Summary

• Add QEff Gemma4 Unified wrappers for text, causal LM, and conditional-generation paths.
• Wire Gemma4 Unified into model replacement, supported architecture, KV-cache, and auto-model flows.
• Add dual-QPC vision/lang export support and an image-text example for Gemma4 Unified.
• Preserve prefill chunking behavior by keeping compile prefill seq_len independent of image prompt length.
• Use vision_config.num_soft_tokens for Gemma4 Unified vision token sizing.

Vision Model Note

Gemma4 Unified changes the vision path from the older heavy vision tower style to a lightweight/encoder-free image embedding path: image patches are embedded through embed_vision, then inserted into the language stream as soft vision tokens. The full model config uses num_soft_tokens=280, so export/runtime now size the retained vision embeddings from that value instead of falling back to older defaults. This matches the model behavior described in the Gemma4 visual guide context (newsletter.maartengrootendorst.com)

Validation

• Verified HF original PyTorch vs QEff modified PyTorch parity.
• Verified QEff PyTorch vs ONNX Runtime parity for the language graph.
• Ran reduced cached-model E2E smoke with QEFF_CTX_LEN=2048, QEFF_PREFILL_SEQ_LEN=128, and QEFF_GENERATION_LEN=64.

cc: @anujgupt-github @quic-hemagnih @tchawada