Add `BlockDiagonalMatrix`

[timweiland]

In a Nutshell

Add a custom linear operator BlockDiagonalMatrix for block diagonal structures.

Detailed Description

• Add BlockDiagonalMatrix along with custom implementations of det, trace, eigvals, etc...
• Add two test cases for it
• Fix bug in eigvals test. Previously this checked the arrays returned by linop and matrix for equality. But the eigenvalues may have a different order. For the comparison to be valid, we need to sort the arrays before comparing them.

[codecov[bot]]

Codecov Report

Merging #794 (5a0843f) into main (384c739) will increase coverage by 0.05%. The diff coverage is 95.94%.

@@            Coverage Diff             @@
##             main     #794      +/-   ##
==========================================
+ Coverage   91.23%   91.28%   +0.05%     
==========================================
  Files         217      218       +1     
  Lines        8133     8207      +74     
  Branches     1027     1059      +32     
==========================================
+ Hits         7420     7492      +72     
- Misses        485      487       +2     
  Partials      228      228              

Impacted Files	Coverage Δ
src/probnum/linops/_block.py	95.83% <95.83%> (ø)
src/probnum/linops/__init__.py	100.00% <100.00%> (ø)
src/probnum/linops/_linear_operator.py	87.17% <0.00%> (+0.17%)	:arrow_up:

[JonathanWenger]

Yes, that's reasonable. If we need non-square blocks in the future that can be refactored. No need ro design too generically at this point.

[timweiland]

Thanks for the reviews! Upon further thought and after discussion with @marvinpfoertner, I decided not to require all blocks to be square. For example, BlockDiagonalMatrix is useful for implementing independent multi-output kernels. When we consider k(X, X), everything is fine and all blocks are square, but k(x, X) (i.e. the cross-covariances between the input and the training data) will have non-square blocks in general. So there is a good use case for it, I think.