Replace magic constant test for linear solve on triangular matrices

[jiahao]

• Provides the Skeel condition numbers condskeel
• Implements the error bounds derived in [JIN 5705](https://github.com/jiahao/jin/blob/master/JIN5705 - Distribution of linear algebra test failures for triangular matrices.ipynb) (nbviewer)

[jiahao]

cc: @staticfloat @mschauer

[staticfloat]

This passes all tests on OSX with OpenBLAS and Accelerate.

[mschauer]

After running the test and subsequently commenting out lines 52, 54, 69, 156, 522, 523 etc.

#        @test_approx_eq apd * inv(capd) eye(n)
...
#        @test_approx_eq det(capd) det(apd)
...
#            @test_approx_eq apd * inv(cpapd) eye(n)
...
#        @test_approx_eq_eps prod(f[:values]) prod(eigvals(a[1:5,1:5]/a[6:10,6:1 
...
#    @test_approx_eq W\v iFv
#    @test_approx_eq W\v iFv
#    @test_approx_eq det(W) det(F)
...
#        @test_approx_eq iWv iFv

the test still fail. This is unrelated to your change but there is no point to proceed further testing on 32 bit with matrices that close to singularity.

[mschauer]

Anyway, I gave it a last try: No specific 32 bit issues with your PR. 👍

[jiahao]

Thanks for cross-checking.

[stevengj]

@jiahao, why were the Integer versions needed, since inv(A) already converts to float?

Note also that abs(inv(A))*abs(A)*abs(x) should be abs(inv(A))*(abs(A)*abs(x)) in order to perform matvec and not matmul operations.

See also the discussion in #17302.

[stevengj]

cc @Sacha0