DEV: remove _arg from public API, add matmul to benchmark ufuncs

[mattip]

Fixes #12682 by removing the private _arg ufunc used for testing from the public API and adding matmul to the list of ufuncs in bench_ufuncs

[seberg]

Looks good, removal of _arg is definitely good. I wonder if we want gufunc in that benchmark, since it requires that the later matrices are aligned properly (which they currently are since they are squared), or whether we should just check for ufuncs with ufunc.signature is None.

[eric-wieser]

Looks fine to me

[seberg]

OK, lets put this in. If we ever add ufuncs that cannot be tested together as a bunch here, we can still split it up into ufuncs and gufuncs.

[seberg]

Are we still keeping commit messages starting with things? BENCH: is a bit nicer then.