Fix stochastic failure in WalkerAlias

[asi1024]

Fix #6045.

[niboshi]

Please write come comments. This workaround is not intuitive.

[niboshi]

How about this? I think it's simpler.

    def sample_xp(self, xp, shape):
        thr_dtype = self.threshold.dtype
        pb64 = xp.random.uniform(0, len(self.threshold), shape)
        index = pb64.astype(numpy.int32)
        left_right = (
            self.threshold[index]
            < (pb64.astype(thr_dtype) - index.astype(thr_dtype)))
        left_right = left_right.astype(numpy.int32)
        return self.values[index * 2 + left_right]

By the way, it seems this problem has been introduced in this commit.
(Apparently, the purpose of the fix was to support ChainerX which does not support implicit dtype promotion.)

[niboshi]

Jenkins, test this please

[chainer-ci]

Jenkins CI test (for commit 2d75429, target branch master) succeeded!

[niboshi]

LGTM