This fixes an issue about mindspore process cannot exit when calling python api op_select_format failed in select kernel steps.
Previously function op_select_format and check_supported raise an exception directly on the tbe_process python side, but we don't deal with the exception, and raise an exeception on c++ side to frontend ME, that will cause some conflict when recycle resource on ME and tbe_process python interpreter.
This changes adding try...catch in function op_select_format and check_supported on the python side, and return the Exception string to c++ side, so that we can raise an exception to frontend ME and ME will deal with resouce clearning and exit.

Signed-off-by: NTing Wang <kathy.wangting@huawei.com>

Signed-off-by: NTing Wang <kathy.wangting@huawei.com>

Signed-off-by: NJonathan Yan <jonathan.yan1@huawei.com>

1.Change dtype of scale to dtype of grad in loss_scale.py;
2.Change dtype of weight_decay to dtype of weight in optimizer.py.

Merge pull request !7 from JoyLvliang/ms-incubator-pynative-add-lenet

Merge pull request !22 from chenhaozhe/enable-variable-acc-for-training-graph

Merge pull request !15 from changzherui/master

!11 Fix bug: Failed to get parameter data in model parallel scenes, when the size of tensor large than 2GB 
Merge pull request !11 from WeibiaoYu/fix-mem-problem

Merge pull request !10 from zhangzheng/SparseApplyFtrlD