Reduce mul! methods related to "banded" matrices

[dkarrasch]

I was encouraged by @blegat to further reduce the number of mul! methods, so here we go. I think we can do this by defining larger "groups" of types, catch any multiplication where they are involved and redirect to its own multiplication function(!), so here it's _mul!. This may be of interest to @dlfivefifty: I grouped together all the "banded" structured matrices (remember we discussed subtyping (some of) them to AbstractTriangular), and called the group BandedMatrix. Currently, this is a simple union, but we may want to consider actually introducing an abstract type and make these subtype it. And we may want to call the "banded matrix-related" _mul! bandedmul!. I can guess you can see where this is going. Does that sound like a good direction or should we avoid any interference with *BandedMatrices.jl?

In the big scheme of things, we can then resolve ambiguities with other "groups"(like wrapped (typically dense) matrices à la triangular, upper Hessenberg) at the larger scale, and ambiguity resolution within the groups happens at the _mul!/bandedmul! scale, so doesn't increase the number of mul! methods.

This PR reduces the number of mul! methods from 88 to 67, but it excludes the triangular part, which is a huge source of methods. I''ll take care of that part in my triangular overhaul #43972.

[dkarrasch]

Let's merge this and discuss naming over at #43972.

[dlfivefifty]

Why not just make them subtypes of AbstractBandedMatrix which then BandedMatrices.jl can use?

(though that’d take work to implement since then we’d lose the LayoutMatrix functionality)