Initial setup for export benchmarking suite and some models

[tugsbayasgalan]

Export team would like to build a benchmarking suite for rising popular models. This diff is the first basic infra we will use to setup the models. Models added:

1. Qwen2
2. Kokoro
3. MiniCPM (only the ConditionalChatTTS part)

Results:

{'environ': {'pytorch_git_version': '555bc8349d67a8f5ccd5325d1a7f46c3619b1769'},
 'metrics': {'errors': {'hf_Qwen2': 'Unsupported input type <class '
                                    "'transformers.cache_utils.DynamicCache'>",
                        'hf_minicpm': 'cannot mutate tensors with frozen '
                                      'storage',
                        'kokoro': '.numpy() is not supported for tensor '
                                  'subclasses.'},
             'success_rate': 0.0},
 'name': 'export_new_models'}

[xuzhao9]

The code LGTM, but it will break our CUDA CI which is 40GB A100. Is this expected?
If we can't shrink the model to single GPU, please add this model to test bypass list like https://github.com/pytorch/benchmark/blob/main/torchbenchmark/models/hf_T5_large/metadata.yaml#L11