[TEST][FLAKY] fix random fail

[FrozenGene]

fix #6309

[jcf94]

Should we set the random seed for python random module & numpy random module?

Nerver mind, I noticed that this only used the random engine inside TVM.

[tqchen]

We should avoid the usage of seed to resolve the test error wrt to randomness. Since setting seed can hide problems in real usecases.

Instead, we should make sure the testcase pass for most randomness generated, perhaps by increasing the error change range in this case

[FrozenGene]

@tqchen I undenstand your concern. However, I have done some experiments and the generated data is random, we can not make sure completely like np.mean(na) < 0.2 always be true and developers will complaint it again. So to avoid this condition, I have decided to do this change, we have still other assert could help us verify (like np.min / np.max / random_fill test case etc) However, if you still think we should incerease, I could help to do it (like increasing the shape we generated and incerease the error range to make us CI pass more robust)

[tqchen]

I do think increase the size and error range should help, and as long as the error happens in low probability(we can calculate the confidence) we should be fine.

[FrozenGene]

Have increased size and error. When to run 100 times, previous code will occur 6 errors, 1000 times will occur 44 errors. New code doesn't have check error. Wish this could let CI be more robust. @tqchen