With #49218 about to hit release, we want to do some cleanup on the models we use for Leo

• Qwen 3 14b has been replaced by Qwen 3 VL 30B on the backend, the label should be updated to match
• Qwen 3 VL makes Gemma redundant, we can remove it and free up the GPUs
• Qwen 3 235B will be replaced by self-hosted Qwen 3 235B VL, the label should be updated to match
• Deepseek 3.1 & GPT-OSS should make Deepseek R1 redundant, we can remove it