Use by default generic lu factorization since it is faster.

[saschatimme]

This starts to use the pure Julia lu factorization instead of the LAPACK version.
Here is a small benchmark for square complex matrices. The LAPACK factorization seems to be more precise but also slower. Especially for n>16 it seems that they start to use blocking techniques
which make things much slower.

Size - generic - blas - difference
4 - 0.508μs - 0.652μs - 1.352e-15
5 - 0.641μs - 0.856μs - 4.131e-16
6 - 0.801μs - 1.031μs - 2.277e-15
7 - 0.974μs - 1.291μs - 1.337e-15
8 - 1.211μs - 1.433μs - 8.709e-15
9 - 1.427μs - 1.752μs - 5.333e-15
10 - 1.669μs - 2.285μs - 3.211e-15
11 - 1.970μs - 2.781μs - 7.021e-15
12 - 2.319μs - 2.955μs - 6.036e-15
13 - 2.748μs - 3.627μs - 7.234e-15
14 - 3.130μs - 3.678μs - 2.297e-14
15 - 3.599μs - 4.265μs - 2.805e-15
16 - 4.091μs - 4.445μs - 1.039e-14
17 - 4.592μs - 7.034μs - 7.356e-15
18 - 5.169μs - 9.022μs - 1.576e-14
19 - 5.840μs - 11.238μs - 2.088e-14
20 - 6.528μs - 13.371μs - 4.375e-15

[codecov]

Codecov Report

Merging #65 into dev will increase coverage by 0.4%.
The diff coverage is 100%.

@@            Coverage Diff            @@
##              dev      #65     +/-   ##
=========================================
+ Coverage   91.71%   92.12%   +0.4%     
=========================================
  Files          38       38             
  Lines        1243     1244      +1     
=========================================
+ Hits         1140     1146      +6     
+ Misses        103       98      -5

Impacted Files	Coverage Δ
src/predictors/base.jl	100% <100%> (ø)	⬆️
src/utilities.jl	76.19% <100%> (+0.38%)	⬆️
src/correctors/newton.jl	100% <100%> (ø)	⬆️
src/solving/solve.jl	98.03% <0%> (-1.97%)	⬇️
src/solving/path_crossing.jl	100% <0%> (+12.24%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update d8cbdce...1558632. Read the comment docs.

[saschatimme]

I wanted to add the old version as well, but always hit a compiler bug similar to #22355, so this is out for now.