Created by: juncaipeng

In order to add ci check for check_grad and upgrade the precision of test from float to double, we modify the op_test.py as follows:

• replace old OpTest with OpTestBase, and use all inputs instead only one input to infer dtype
• add OpTestInt8, which is the parent class of the op test with int8 precision
• add OpTestFp16, which is the parent class of the op test with float16 precision. OpTestFp16 demands the test to call check_grad at least once.
• add OpTest, which is the parent class of the op test with float32/64 precision. OpTest demands the test to call check_grad with float64 precision at least once.

TODO:

• add restrictions to numeric_grad_delta, max_relative_error for check_grad with float64 precision
• add check for op_whit_list.py in check_api_approve.sh
• forbid op test to inherit OpTestBase by add check in check_api_approve.sh

TODO: 如果从fp32升级到fp64检查，fp32检查可以不需要了。 简单的改一下单测中fp16继承关系，使得存量问列表最小集

NO_NEED_FP64_CHECK_GRAD_OP_LIST：真正的白名单，应该是那些需要被豁免检查的op或者case NEED_TO_FIX_OP_LIST：是存量问题