speculative : fix seg fault in certain cases

[ggerganov]

Fixes a crash when using tree-based speculative decoding with greedy sampling:

make -j llama-speculative && ./bin/llama-speculative -m ../models/qwen2.5-32b-coder-instruct/ggml-model-q4_0.gguf -md ../models/qwen2.5-0.5b-coder-instruct/ggml-model-q4_0.gguf --ctx-size 0 -ub 4096 -b 4096 -ngl 99 -ngld 99 -fa --draft-max 8 --draft-min 0 --draft-p-min 0.75 -c 4096 --color -p "Write a quicksort algorithm" -np 4 --top-k 1 -s 1