Add specialized methods for complex-real BLAS multiplication.

[andreasnoack]

This implements the method proposed by @stevengj for multiplying a complex matrix with a real matrix or vector. On my computer this gives around a 12x speed up for two 1000x1000 matrices over calling the generic algorithm which is what happens now and it is a double speed up over converting the real array to complex.

The only methods where (I can see that) this is feasible are A_mul_B(Matrix{Complex},MatOrVec{Real}), A_mul_Bt(Matrix{Complex},Matrix}) and by transposing the result also A_mul_Bt(Matrix{Real},Matrix{Complex}) and At_mul_Bt(Matrix{Real},Matrix{Complex}). A_mul_B(Matrix{Real},Vector{Complex}) could be written in terms of two dgemv call with stride=2, but in my benchmarks this is not faster than the generic algorithm.

Finally, I tried to write methods for dot but it was not faster than the Julia version. Not surprisingly, it appears that it is mainly for level 3 operations that BLAS gives big speed gains.

[andreasnoack]

I have removed the two versions that transposed the result. It is probably better just to convert the input matrix to complex for these two. For larger matrices this should be the approach except for A_mul_B(Matrix{Complex},MatOrVec{Real}) and A_mul_Bt(Matrix{Complex},Matrix}) which can use the trick. Related to #3239

[andreasnoack]

I'll merge this tonight if I hear no complains. Note that this pull request also adds exclamation marks to the mutating multiplication functions. This is noted in the commit message, but not in the comments.

[carlobaldassi]

I'll merge this tonight if I hear no complains.

Bump. (12 days later, no complains were heard :) )

[andreasnoack]

I know, but during that day I realised that @stevengj also worked to matmul.jl and decided to wait for him to merge. I think many of the headers in the file should be AbstractMatrix after the recent development in the type hierarchy for arrays, but that can be in another pull request. I'll merge this now.