Test statistics of initializers

[toslunar]

Test scale arg of chainer.initializers.

This PR adds statistical tests and fixes bias of sign of Orthogonal initializer.

[toslunar]

@sunaemon Could you review the math part of this PR?

[toslunar]

This PR can be backported because the change to Orthogonal is moved to #5615.

[mitmul]

Will you make the test case for the orthogonal initializer stricter by removing abs over the sample values after this PR is merged and backported for v5? Or isn't it necessaliry be such a strict condition for testing the orthogonal initializer?

[toslunar]

#5615 makes the test stricter by removing abs.

[sunaemon]

@toslunar
The statistical tests for orthogonal initializer seem verifies a necessary condition!

Sorry for the delay.

[mitmul]

Jenkins, test this please

[chainer-ci]

Jenkins CI test (for commit e792034, target branch master) failed with status FAILURE.
(For contributors, please wait until the reviewer confirms the details of the error.)

[toslunar]

A test (with alpha=0.05 and repeat_with_success_at_least(5, 3)) fails with probability 0.001158125. This seems too large to have parameterized tests.

[toslunar]

Jenkins, test this please.

[chainer-ci]

Jenkins CI test (for commit 44b63fc, target branch master) failed with status FAILURE.
(For contributors, please wait until the reviewer confirms the details of the error.)

[mitmul]

Jenkins, test this please

[chainer-ci]

Jenkins CI test (for commit 44b63fc, target branch master) succeeded!

[okuta]

@mitmul Test was passed.