model: add support for clef decision model (text-only)

[ngxson]

Overview

Follow-up #29818

Pre-quantized models:

llama serve -hf ggml-org/Clef-GGUF
llama serve -hf ggml-org/Clef-Flash-GGUF

Additional information

So far, this model is significantly more complicated and requires a new API call to llama_batch_ext

Unlike traditional models that calculating score in a rerank-similar manner, clef selectively run calc mean over a span of input tokens, example:

... INSTRUCTION:  Is the customer angry ?  \nALLOWED OPTIONS:\nOPTION 1:  {"option_id":"true"}  \nOPTION 2:  {"option_id":"false"}  ...
                  └──── 41..46 ────┘                               └───── 52..58 ─────┘            └────── 62..69 ──────┘
                    question span                                      option 0 span                   option 1 span

A new API llama_batch_ext_set_decision_order(bath, idx, order) is added to mark the order span, example:

order	meaning
0	ordinary text (default)
1, 2, 3	question text, of kind noul / choice / score
4	option text

Example of 2 questions:

text  [Q noul] text [opt] text [opt] text  [Q choice] text [opt] text [opt] text [opt] text
 0     1 1 1    0   4 4    0   4 4    0     2 2 2      0   4 4    0   4 4    0   4      0

Known limitations:

• No vision support yet, requires llama: support both embd + raw tokens in batch #29622 to be merged
• Single seq per batch (ref: TAG_SPLIT_QGATE_QWEN)

TODO:

• make gguf
• clean up unused code
• maybe optimize graph, make topology static

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: yes