We collect graph and dataset graph in begin stage before,
If there compile graph fail in GPU, we also collect graph
and dataset graph to summary dir, it will confuse user.

So we collect graph and dataset graph in step end stage now,
If there compile graph fail, we will not collect graph and dataset
graph.

Merge pull request !3651 from jiangzhiwen/dataset/change_num_samples

Merge pull request !3625 from hewei/optimize_tensor_data

Merge pull request !3644 from liyong126/fix_mindrecord_bug

Merge pull request !3638 from geekun/yjk_master

Merge pull request !3643 from huanghui/communication-op-fusion

Merge pull request !3489 from lvchangquan/master

Merge pull request !3594 from wangqiuliang/fix-batchnorm-under-mix-precision-in-pynative

Signed-off-by: GuoMengHao <guomenghao@huawei.com>

Merge pull request !3630 from laiyongqiang/refnode_input_fix

Merge pull request !3635 from panfengfeng/fix_dataset_train_gil_of_gpu_master

Merge pull request !3627 from guozhijian/fix_device_occupied_tdt_hung

Merge pull request !3624 from yanghaitao/yht_generator_timeout

Merge pull request !3574 from hewei/rename_user_data_func

Merge pull request !3637 from thlinh/dev_Jul28_lower_checking_threshold

Replace std::vector<T> with std::unique_ptr<T[]> for tensor data storage,
it prevent unintended data initialization when data is lazy allocated.

Merge pull request !3603 from jiangzhiwen/fix_column_names_exceeded

Merge pull request !3620 from 张学同/to_merge

Merge pull request !3270 from thlinh/dev_Jul17_removeSelect

Merge pull request !3611 from ms_yan/vgg_repair

Merge pull request !3556 from jinyaohui/master

Merge pull request !3612 from changzherui/mod_case

Merge pull request !3563 from zhouyuanshen/master

Merge pull request !3615 from XunDeng/pp_poc_v3

Merge pull request !3614 from Giancarlo/update_convert_sw

Merge pull request !3578 from meixiaowei/master

Merge pull request !3606 from vlne-v1/amp_doc