Fix Intel vs GNU ABI when building against MKL

For functions that return complex values, there are two strategies how values
can be returned. Either via a value on the stack (aka return value) or via an
implicit first argument (aka output argument). The former is called GNU ABI and
the latter Intel ABI by the Intel MKL. There are two variants of the Intel MKL
and when calling functions like zdotu that return a complex value it does make a
difference which one the application is linked to.

Unfortunately, there is no way to detect this at runtime and other libraries
that are linked into the application might have other preferences with regards
to which MKL interface they want to use. With this commit we make the code ABI
agnostic by simply using the CBLAS API which always uses an output argument.

https://www.intel.com/content/www/us/en/docs/onemkl/developer-guide-linux/2024-1/call-blas-funcs-return-complex-values-in-c-code.html

c++/nda/blas/interface/cxx_interface.cpp

@@ -83,11 +83,21 @@ namespace nda::blas::f77 {
 
   double dot(int M, const double *x, int incx, const double *Y, int incy) { return F77_ddot(&M, x, &incx, Y, &incy); }
   dcomplex dot(int M, const dcomplex *x, int incx, const dcomplex *Y, int incy) {
+#ifdef NDA_USE_MKL
+    MKL_Complex16 result;
+    cblas_zdotu_sub(M, mklcplx(x), incx, mklcplx(Y), incy, &result);
+#else
     auto result = F77_zdotu(&M, blacplx(x), &incx, blacplx(Y), &incy);
+#endif
     return dcomplex{result.real, result.imag};
   }
   dcomplex dotc(int M, const dcomplex *x, int incx, const dcomplex *Y, int incy) {
+#ifdef NDA_USE_MKL
+    MKL_Complex16 result;
+    cblas_zdotc_sub(M, mklcplx(x), incx, mklcplx(Y), incy, &result);
+#else
     auto result = F77_zdotc(&M, blacplx(x), &incx, blacplx(Y), &incy);
+#endif
     return dcomplex{result.real, result.imag};
   }