T5gemma2 model support

[timkaye11]

Adds T5Gemma-2 Model Support for embedding and text generation

relies on ajroetker/hugot#4

Embeddings (pkg/termite/lib/embeddings/):

• t5gemma2.go (+606 lines) - T5Gemma2 embedder with text & image embedding support
• t5gemma2_stub.go (+97 lines) - Non-ONNX build stub
• hidden_states.go (+50 lines) - Hidden states extraction for advanced use cases

Generation/Rewriting (pkg/termite/lib/seq2seq/):

• t5gemma2.go (+588 lines) - T5Gemma2 generator with multimodal seq2seq support
• t5gemma2_stub.go (+143 lines) - Non-ONNX build stub

Export Scripts

• export_t5gemma2.py (+1,143 lines)
  • Self-contained ONNX export script with:
  • Encoder, decoder-init, decoder, and vision encoder export
• export_model_to_registry.py (+211 lines) - Enhanced model registry export

Documentation & Tests**

• e2e/t5gemma2_test.go (+380 lines) - E2E tests for embedder and generator

[dovinmu]

Claude's assessment of this branch for VLM (scoping for v2 of Honeycomb pipeline):

The branch is incomplete. The vision encoder works, but image embeddings aren't injected into the decoder — so image→caption doesn't actually work.