dispatch.jl: remove LinearAlgebra.mul! overloads

[odow]

This one is a little more controversial. Originally added by @blegat in #16

At first glance, the problem with this is that we're missing a bunch of methods. Triangular, Symmetric, etc.

I started writing a replacement that would loop over the types in LinearAlgebra. But this just ended up creating a lot of ambiguities because SparseArrays does something similar:
https://github.com/JuliaSparse/SparseArrays.jl/blob/311b4b4130d9f28a6b5eb55cb7c818e4f7858719/src/linalg.jl#L30-L56

So then I stepped back and asked, why are we even implementing these methods? MutableArithmetics has plenty of support for operate_to!(C, *, A, B), and trying to intercept all of these weird array types based on the element type is ripe for problems.

I couldn't really find examples of packages where we call this (it's quite hard to search for mul!), so this might lead to a performance regression in some user-code, but arguably they should be using the proper MA API instead of trusting this weird fallback in LinearAlgebra?

[odow]

These are also candidates to remove: #181

[blegat]

When the user calls A * B outside the macro, it ends up calling LinearAlgebra.mul! which either fails because it incorrectly preallocate the resulting vector for JuMP types (see #47) or is slow because it does not exploit mutability so JuMP needs this

[odow]

When the user calls A * B outside the macro

Yes. I don't see why we should intercept calls outside the macros.

I guess if this is a breaking change, we can not merge it, but for MutableArithmetics 2.0, I'm not in favor.

[odow]

Our other problem is that I can remove all of these methods and the tests still pass.

[odow]

I guess we needed this because the old rewrite macro didn't rewrite every expression. It would often leave complicated expressions as-is, so x' * (A * x) might leave A * x alone, and that would need the fallback.

[odow]

I'm adding some tests to JuMP that demonstrates why we need these: jump-dev/JuMP.jl#3132

[odow]

Okay. The new JuMP tests are now failing with the removal of these methods.

[blegat]

It is suprising that the JuMP tests were not failing indeed, it's good to have jump-dev/JuMP.jl#3132