Version 5.1

[Jutho]

This contains two commits, which should not be squashed.

1. this changes the strided implementation so as to always specialise, getting rid of small allocations and some overhead for small permutations
2. this abolishes the get_extension trick in the extensions which started to break on Julia 1.11.1, and simplifies the rrules for the tensor operations so that also other packages (i.e. TensorKit) should not rely on that same broken trick to overload the rrules

Along the way, it also fixes the blas_contract overload in the bumper extension so that it actually gets used.

[Jutho]

Failures on x86 with Julia 1(.11) are CUDA related.

[maleadt]

I take it this fixes the LoadError: StridedViewsCUDAExt not found observed in JuliaGPU/CUDA.jl#2525?

[Jutho]

Yes, that was another consequence of the "extension loading during precompilation" change in Julia 1.11.1

[maleadt]

Thanks! Do you plan to put this in a release anytime soon?

[Jutho]

Yes, I want to test one more thing this afternoon, and then I will tag the release.