skip_tokenizer_init=True crashes google/gemma-3-27b-it

workaround is to set it to false for gemma models. Filed issue upstream.

https://github.com/vllm-project/vllm/issues/31123

once fixed we can remove the skip init force in RL