Add support of framework Accelerate for MacOS

[foxtran]

This patch masks old buggy implementation of BLAS/LAPACK from Apple Accelerate with fixed version via aliasing. New implementation is available since MacOS 13.3, and has suffix $NEWLAPACK for LP64 mode, ILP64 mode has syffix $NEWLAPACK$ILP64. However, new implementation still has incompatible interface for cdotc, cdotu, zdotc, zdotu. Fortunately, xtb does not use them.

I've masked all used BLAS/LAPACK routines which are used at the current moment. Better solution is to provide explicit interfaces/masking to BLAS/LAPACK in source code.

Closes #1319

[awvwgk]

I am not too happy with this change because it adds a long list of flags globally. Would it be possible to implement this on the source level by overriding the BLAS / LAPACK interfaces with an extra bind(c, name="_dxyz$NEWLAPACK")?

[foxtran]

Would it be possible to implement this on the source level by overriding the BLAS / LAPACK interfaces with an extra bind(c, name="_dxyz$NEWLAPACK")?

Actually, I have a library for doing this... However, it requires to update all deps of xtb and this is really painful.

[foxtran]

And one has to instrument whole xtb source code with these interfaces. For example, here:

xtb/src/bfgs.f90

Line 48 in 6d44803

real(wp) :: ddot, thrs, scal, damp, dampO,dampD,thr

and here:

xtb/src/bfgs.f90

Lines 144 to 147 in 6d44803

	dds = ddot(nat3,dx,1,dx,1)
	if(dds > thrs) then
	ddtd = ddot(nat3,tvec,1,dx,1)/dds

ddot is used without mctc interface. So, it is a big work.

For MRCC, I solved this problem mainly with patched old compiler, but it can not understand xtb source code, unfortunately.

[philipturner]

While trying to inject a custom DSYEVD or SSYEVD kernel, it was troublesome to tabulate all possible linear algebra functions used in xTB, and then create a wrapper library that calls standard BLAS for every call except xSYEVD. If we did have a clear way to selectively migrate specific BLAS functions to custom libraries, that would help a little bit with future performance improvements.

https://github.com/philipturner/nanotech-code-archive/blob/main/swift-xtb/Experiments/PerformancePrediction/Trampoline.swift

However, at some point, for example to switch between FP64 and FP32 based on convergence criteria, the client would have to recompile xTB from source. This is just food for thought.

[foxtran]

Hmm.. Something is broken...

[foxtran]

@thfroitzheim, ping me next time when you want to update code. Something goes wrong.

[thfroitzheim]

We wanted to check if your addition of accelarate would fix the current problem on macos in #1396. But it doesn't seem to be the case. Everything else still seems to work here.

[foxtran]

Oh.. I see.. Can fix it at evening.