Fortran is also optimized and debugged over decades of testing.
Yeah those algorithms are numerically tested, haven't changed in years, and really fast. You can certainly get to convenient matrix manipulation in C++, but then your compile times will increase significantly.
You can have a fast implementation and fast compilation time with DLang
Performance might be worse in c++
Can the fortran implementations do auto-vectorization on simd hardware? or map to gpu using cuda-like library?
Yes, in fact all the original research for those things was done in Fortran compilers.
Fortran is a fully-vectorized and parallelized language at all levels, from instructions to shared and distributed parallelism (corresponding to one-sided MPI communications). This has been the case for at least 3 decades if not more. There are mature GPU support for Fortran. NVIDIA and PGI are currently implementing the standard parallel features of Fortran via GPUs.