Fix flan-t5 models load fail with llama-server

[kylo5aby]

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

Fixes: #8990

[fairydreaming]

I'm not sure if there is a point in doing this, since T5 models require llama_encode() call and preparation of a custom input for llama_decode() (with decoder start tokens). This is not implemented in the current llama-server implementation, so even if the model loads it won't work correctly. But if you are willing to extend your PR by implementing the comprehensive support for T5 models in llama-server then you have my blessings.

[fairydreaming]

It looks like T5 model loading in llama-server is already fixed by #8997

[ggerganov]

Yup, this should fix the loading, but llama_encode() support needs more work