[misc] fix: Use AutoModelForImageTextToText instead of AutoModelForVision2Seq

[kaln27]

What does this PR do?

• According to Deprecate AutoModelForVision2Seq huggingface/transformers#38900, the AutoModelForVision2Seq will be deprecated soon. So verl need to use AutoModelForImageTextToText instead.
• Also I found some model doesn't have 'generation_config.json` file, but the code assume it has. It will cause an error. So I check whether it is existed first (Already fix by other PR) .
• I also create a PR Fix errors when use verl to train GLM4.1v model huggingface/transformers#39199 in transformers to support GLM4.1v training in verl.
• Add GLM4.1v flops counter function. (Already done bu other PR)

[gemini-code-assist]

Code Review

This pull request updates the checkpoint saving logic to handle cases where generation_config.json is missing and adds support for glm4v in the FLOPs counter.

My main feedback is on the fsdp_checkpoint_manager.py file. The current implementation for checking the existence of generation_config.json introduces a potential race condition. I've suggested a more robust try-except pattern to prevent potential crashes.

The changes in flops_counter.py look good.

I noticed the PR title mentions replacing AutoModelForVision2Seq with AutoModelFotImageTextToText, but this change doesn't seem to be included in the diffs. You might want to double-check if that was intended for this PR.

[CLAassistant]

All committers have signed the CLA.

[kaln27]

@eric-haibin-lin Hi eric. Could you review this PR? Thanks a lot.