xtb version 6.4.0

Feature release with new GPU acceleration, improved ALPB parameters, better optimizer settings and options, and general performance improvements for GFN1-xTB and GFN2-xTB.
Thanks to Sebastian Spicher (@sespic), Marcel Müller (@marcelmbn), Hagen Neugebauer (@haneug), Thomas Rose (@Thomas3R), Marcel Stahn (@MtoLStoN), Julius Stückrath (@jbstueckrath), Patrick Atkinson (@patrickatkinson), Cyrille Lavigne (@clavigne) and Igor Gerasimov (@FulgurIgor) for contributing to this version.

• Bugfix: Basename detection failed for paths with path delimiter (#362) 🪲
• Bugfix: FOD flag (--fod) did not set electronic temperature correctly (#382) 🪲
• Bugfix: Sign error in unbiased hessian calculation (#388) 🪲
• Bugfix: IR intensities were calculated incorrectly from dipole gradients (#395) 🪲
• Bugfix: Fix out-of-bounds access for basis set in IPEA-xTB calculations (#399) 🪲
• Bugfix: Commandline interface for sequential build was inconsistent with OpenMP build (#421) 🪲
• Bugfix: Point charge embedding was not initialized correctly and skipped embedding calculations (#422) 🪲
• Bugfix: Negative charges were detected as command line flags (#423) 🪲
• Improved 2D to 3D conversion with GFN-FF (#351)
• GPU version of xtb is available with Nvidia HPC SDK (#342, #352, #407, #420)
• Backports of performance improvements from GPU routines to CPU implementations (#425)
• Allow adjusting of spring exponents in distance constraints (#357)
• Allow scaling of dispersion energies / gradients for the GFN-FF (#359)
• Improved ALPB implicit solvation parameters (#372, #387)
• Turbomole driver now attempts to cache results via gradient file (#377)
• Adjusted optimization threshold for level “lax” for Turbomole DCOSMO-RS driver (#383)
• Allow reading static RMSD bias from file (#390)
• Use averaged convergence scheme in L-ANCopt and RF-ANCopt to deal with noisy energy surfaces (#394)
• Refactored hessian implementation to stabilize build and allow higher compiler optimization (#417)