Broadcasting rather than BLAS 1 calls?

[haampie]

There are some instances where BLAS calls can be replaced by broadcasting.

In CG and Chebyshev iteration there is an axpby which is currently written as scale and axpy (since axpby is not part of Julia 0.6). This looks wasteful:

scale!(cheb.u, β)
@blas! cheb.u += one(T) * cheb.c

could be replaced by

@. cheb.u = cheb.c + β * cheb.u

Is there any reason not to do this? And while we're at it, can't we replace axpy's with broadcasting in general?

cc'ing @andreasnoack because we probably discussed this earlier.

[andreasnoack]

I think that should work and probably faster since it is fused into a single traversal of the elements. As a matter of style, I prefer explicit dots over the @. macro.