One more relaxation of test bounds in linalg.

[mschauer]

(amends #5691, it still failed with 32 bit.)

[stevengj]

It just seems like asking for trouble do linear-algebra tests with a random 10x10 matrix (a = rand(n,n)). There is some probability that the matrix will be badly conditioned enough to fail any test with any reasonable error threshold.

Why not just use a deterministic matrix instead? e.g. a = [exp(cos(i + 2*sin(j)))*(i-j)+sin(i*j) for i = 1:n, j=1:n] is not too badly conditioned for n=10.

[mschauer]

Part of the problem is that when the magic numbers are chosen travis does not test on 32 bit, see travis-ci/travis-ci/issues/986, part of the problem is #5549 .

[jiahao]

Please help cross-reference #5605 on such changes.

[jiahao]

I'll accept this change but all this magic number nonsense just needs to go away.