`test_math` is flaky

This is part of https://github.com/chainer/chainer/issues/6903 effort.

• Failure example: https://ci.preferred.jp/chainer.py37/2259/
• Failure test: chainerx_tests/unit_tests/routines_tests/test_math.py
• Failure test case: test_Tan_param_17_{contiguous=None, in_dtypes=('float16',), out_dtype='float16', input=2, shape=(2, 3)}[native:0]
• Tolerance:

E           Not equal to tolerance rtol=0.001, atol=0.001
E           
E           Mismatch: 100%
E           Max absolute difference: 0.00203618
E           Max relative difference: 0.00200944

[imos]

It seems the test is still flaky in the master branch.

• Failure example: https://ci.preferred.jp/chainer.py37/2762/
• Failure test: chainerx_tests/unit_tests/routines_tests/test_math.py
• Failure test case: test_LogSumExp_param_1_{keepdims=False}_param_1_{shape=(2,), axis=-1}_param_0_{in_dtypes=('float16',), out_dtype='float16'}[cuda:0]
• Tolerance:

E       Not equal to tolerance rtol=0.001, atol=0.001
E       
E       Mismatch: 100%
E       Max absolute difference: 0.00116
E       Max relative difference: 0.011566

[asi1024]

• Failure example: https://travis-ci.org/chainer/chainer/jobs/522915728#L3603
• Failure test case: test_Abs_param_24_{shape=(), contiguous='C', out_dtype='float32', input='random', in_dtypes=('float32',)}[native:0]