Efficiency improvements, add multi-threaded matvec

[ChrisRackauckas]

Why not use Base's sparse matvec?

[coveralls]

Coverage decreased (-1.2%) to 89.944% when pulling 46e356ac817d26bd644a5450214920b02176b802 on mem into db32186b8ca2131bf1c8fe7d159b9f32ac3c8194 on master.

[ranjanan]

I use that by default, but since there are a lot of matvecs in the package, I may well get speedups on large matrices.

As for the implementation, I could have just copied Base's implementation because it's more generic. I just wanted to play with this simplest form first and run benchmarks to see if it works.

[ranjanan]

I also had plans to add a multi-threaded spmatmul but it looks like I'd have to modify the implementation a little to thread it.