Add support T5 (encode-decoder) models at API level and server

Ciao Andrei,
following the latest changes on llama.cpp repo
[https://github.com/ggerganov/llama.cpp/issues/5763](https://github.com/ggerganov/llama.cpp/issues/5763)

can you tell me if this is going to be art of your amazing repo?
In case it is not, can you provide some instructions to replicate at low Level API the same steps done by lama-cli?

we can start using amazing models like LaMini-Flan-T5-248M with llama-cpp-python.

thanks