remove *(A::Matrix, B::HermOrSym) = A*full(B)

- this dispatches to BLAS instead
- seems a tiny bit faster
- allocates half of the memory
- main reason is to be consistent with the opposite method
  e.g. *(A::HermOrSym, B::Matrix) which does no longer have this
  fallback method with a full call, but dispatches to
  BLAS.symv!/BLAS.hemv! directly

only call full on second argument for *(A::HermOrSym, B::HermOrSym)

Author: fredrikekre

base/linalg/symmetric.jl

@@ -299,8 +299,7 @@ A_mul_B!(C::StridedMatrix{T}, A::Hermitian{T,<:StridedMatrix}, B::StridedMatrix{
 A_mul_B!(C::StridedMatrix{T}, A::StridedMatrix{T}, B::Hermitian{T,<:StridedMatrix}) where {T<:BlasComplex} =
     BLAS.hemm!('R', B.uplo, one(T), B.data, A, zero(T), C)
 
-*(A::HermOrSym, B::HermOrSym) = full(A)*full(B)
-*(A::StridedMatrix, B::HermOrSym) = A*full(B)
+*(A::HermOrSym, B::HermOrSym) = A*full(B)
 
 for T in (:Symmetric, :Hermitian), op in (:+, :-, :*, :/)
     # Deal with an ambiguous case