support for common hf multimodel

[n1ck-guo]

[wenhuach21]

As discussed: 1 Configure the template using a file to eliminate the need for additional coding. 2 move the code from example to autoround 3 Refine the cache code in AutoRound to improve generalizability for new models with train_batch_size=1.

[wenhuach21]

4 refine all the examples of vlm

[wenhuach21]

5 also support it in auto_round cmd