Don't create temporaries in real eigfact.

[andreasnoack]

@stevengj The reason for the allocation was much simpler than expected. When creating the eigenvector matrix from the LAPACK output, the columns were fetched Matlab-like and therefore a lot of temporaries were created. By avoiding the temporaries, I get

julia> @time eig(randn(1000,1000));
  2.206440 seconds (115 allocations: 38.461 MB, 0.29% gc time)

As noted in #16751, the many extra allocations have a modest effect on the timings, though. The eigenvalue computation completely dominates.

Fixes #16751

[stevengj]

Thanks for looking into this. I figured the O(n^3) LAPACK part would dominate, but it is still nice to see not so many temporaries created. This is also helpful in tracking down allocations in user code, by avoiding a lot of spurious allocations in base.

[ViralBShah]

This is 5x faster for eig of small matrices. I tested 100x100. Perhaps deserves a mention in NEWS.