Basic matrixcompare functionality

[Andlon]

Note: this is not ready for merging. I'm posting it here for some initial feedback.
Update: This is in principle finalized and ready for merging (barring minor issues, see below).

This PR provides matrixcompare functionality for nalgebra. matrixcompare is a crate which aims to provide generic functionality for debug comparisons of matrices. A consumer crate like nalgebra only needs to define how many rows/columns a matrix has and how to fetch coefficients. Macros like assert_matrix_eq! then makes it easy to compare matrices with different kinds of comparators (absolute tolerance, relative tolerance etc.), which significantly simplifies debugging when writing unit tests involving linear algebra code.

matrixcompare is the result of extracting, rewriting and generalizing similar functionality found in the rulinalg crate (and also originally contributed by me).

Note that matrixcompare is currently not published, and is also not complete. Obviously, this will have to be remedied before this PR can be merged. I open it now so that I can perhaps have some initial feedback (including CI runs).

Below is an excerpt from one of the new tests.

let a = MatrixMN::<_, U4, U5>::from_row_slice(&[
        1210, 1320, 1430, 1540, 1650,
        2310, 2420, 2530, 2640, 2750,
        3410, 3520, 3630, 3740, 3850,
        4510, 4620, -4730, 4840, 4950,
]);

let b = MatrixMN::<_, U4, U5>::from_row_slice(&[
        1210, 1320, 1430, 1540, 1650,
        2310, 2420, 2530, 2640, 2750,
        3410, 3520, 3630, 3740, 3850,
        4510, 4620, 4730, 4840, 4950,
]);

assert_matrix_eq!(a, b);

And this is the output (format subject to change) at the time of writing:

Matrices X and Y have 1 mismatched element pairs.
The mismatched elements are listed below, in the format
(row, col): x = X[[row, col]], y = Y[[row, col]].

 (3, 2): x = -4730, y = 4730.

Comparison criterion: exact equality x == y.
Please see the documentation for ways to compare matrices approximately.

[sebcrozet]

Thanks, that looks very useful!

Why is the message method called panic_message at some places? I guess this does not panic, so would it make more sense to call this something like failure_message?

Or perhaps compare_matrices could return a Result<(), MatrixComparisonFailure>, where MatrixComparisonFailure would have its Match variant removed, and would implement Display and Error?

[Andlon]

Thanks for having a look! You're right, panic_message is a terrible name. It's somewhat of an old relic, in which it was used exclusively to panic inside of a macro. But I somehow didn't change the name as I was redesigning the functionality. I like your suggestions, I will try to apply them next chance I get!

EDIT: Also thanks for your suggestion about copy instead of clone. I somehow switched around the name in the commit messages though, oops. Will fix this later too.

[Andlon]

@sebcrozet: I finally managed to revisit this PR. I made the changes you suggested earlier, and matrixcompare has now been released on crates.io. I've updated this PR to use the updated version.

matrixcompare also supports sparse matrices (and dense-sparse comparisons of course), which is why I'm trying to get this across the finishing line now. It makes it much easier to write tests for my upcoming sparse contribution. This PR only includes support for nalgebra dense matrices, however. I will include sparse support in the upcoming CSR/CSC redesign.

The matrixcompare crate is organized in terms of the core traits, that reside in matrixcompare-core, and further re-exported in the matrixcompare crate. The idea here is that libraries like nalgebra need only rely on matrixcompare-core (which has no dependencies and is expected to need updates very rarely, reducing dependency breakage), whereas consumers (e.g. nalgebra's own test suite) rely on whatever version of matrixcompare that they want to in their tests.

I'm not sure how to really do it with features etc. I added a "compare" feature which would toggle the matrixcompare-core dependency. I cannot call it "matrixcompare", because matrixcompare is a dev-dependency already, which causes a conflict. Also, unfortunately I cannot toggle this feature for our tests by default, but I'm sure you're aware of this particular problem.

I see that the CI fails for no-std tests. Not exactly sure how things fit together here, I have basically almost zero experience with no-std.

[sebcrozet]

Thank you for this PR! The CI failure when targeting no-std is caused by the matrixcompare dev-dependency which itself depends on num-traits which has its std dependency enabled. Unfortunately, cargo will unify enabled features for all dependencies, including dev-dependencies, when building the crate. So if num-traits/std is enabled on one dev-dependency, it will be enabled for all builds of nalgebra itself. This should be fixed by Andlon/matrixcompare#2

[sebcrozet]

Thanks!