* save_predict_result

* save_predict_result

* fix model name

* fix: bug when distillation

* modify some default hyperparams to adapt to fine-tune downstream tasks

1. unset EMA because of the relatively small size of most downstream dataset;
2. use mean and std of IMN.

1. fix bias_attr to False for conv of PatchEmbed;
2. support return_tokens_mean for Head of CLIP;
3. support remove_cls_token_in_forward for CLIP;
4. support head_init_scale argument for ViT backbone;
5. support get_num_layers() and no_weight_decay() for ViT backbone.

compatible with paddle 2.5 and older version.
ref: https://github.com/PaddlePaddle/PaddleClas/pull/2798

there are diffs that step()+update() and minimize().
this will be fixed in https://github.com/PaddlePaddle/Paddle/pull/53773.