server: use pos_next instead of n_tokens for m-rope

[am17an]

Overview

Repro: ./build/bin/llama-cli -hf ggml-org/Qwen2.5-VL-7B-Instruct-GGUF:Q8_0 -hfd Qwen/Qwen2.5-0.5B-Instruct-GGUF:Q8_0 -p "what is in this image" --image photo.jpeg, fails in init_batch

init: the tokens of sequence 0 in the input batch have inconsistent sequence positions:                                                                                               
   - the last position stored in the memory module of the context (i.e. the KV cache) for sequence 0 is X = 85                                                                          
   - the tokens for sequence 0 in the input batch have a starting position of Y = 85                                                                                                    
   for M-RoPE, it is required that the position satisfies: X < Y    

seq_rm expects a position and with M-RoPE n_tokens != n_pos so it this bug triggers.

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, AI identified this bug, I repro'ed it and fixed it

[ngxson]

I didn't test it but seems to be a valid fix, CC @ggerganov if you can have a quick look

[ggerganov]

Thanks, good catch.