Loop over diagind in diag for banded matrices

[jishnub]

This improves performance in obtaining the zero bands for banded matrices. This is mainly because of @inbounds annotations.

Using

D = Diagonal(1:6000);
B = Bidiagonal(1:6000, 1:5999, :U);

Operation	master	PR
@btime diag($D, 2);	2.832 μs	1.925 μs
@btime diag($B, 2);	2.283 μs	1.937 μs

The performance difference in Tridiagonal and SymTridiagonal seems pretty negligible, for reasons that I don't fully understand. Perhaps the indexing is a more complicated operation for these, so that the bounds-checking isn't the dominant contributor.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 91.89%. Comparing base (b464203) to head (5de9d05).
Report is 2 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #1211   +/-   ##
=======================================
  Coverage   91.88%   91.89%           
=======================================
  Files          34       34           
  Lines       15374    15378    +4     
=======================================
+ Hits        14127    14131    +4     
  Misses       1247     1247           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.