vocab : reduce debug logs about non-EOG control tokens

[ggerganov]

• Reduce logs for unused control tokens not being marked as EOG
• Append control attribute instead of assigning

Some models have many unused tokens, which leads to a lot of debug logs. F.ex:

./bin/llama-server -hf ggml-org/embeddinggemma-300M-GGUF --embeddings -lv 5

...

0.00.960.046 D init_tokenizer: initializing tokenizer for type 1
0.00.966.704 D load: control token: 262141 '<unused6239>' is not marked as EOG
0.00.966.706 D load: control token: 262140 '<unused6238>' is not marked as EOG
0.00.966.706 D load: control token: 262137 '<unused6235>' is not marked as EOG
0.00.966.707 D load: control token: 262136 '<unused6234>' is not marked as EOG
0.00.966.707 D load: control token: 262135 '<unused6233>' is not marked as EOG
0.00.966.707 D load: control token: 262134 '<unused6232>' is not marked as EOG
...

[CISC]

Just an aside...

In #18500 it was discovered that f.ex. jina-v2-en has lots of [unusedX] tokens that are marked as NORMAL, but were treated as (but not set to) CONTROL due to the [] name.

AutoTokenizer does not treat them as control tokens either, so the new behavior after that PR is the correct one.