GPU-friendly truncation implementations

[lkdvos]

This is an attempt to get rid of the scalar-indexing oriented approach, and instead do more global operations.
Definitely still WIP, and on CPU there are definitely various optimizations that can be applied if needed.
I do wonder about the performance a bit, as I would actually expect that for a large number of sectors this might just be faster.

Some possible optimizations:

• for UniqueFusion, finding the nth value is simply partialsortperm(values, n; by, rev), avoiding the need to allocate the full permutation vector
• for CPU, cumsum + findlast can be replaced by a loop to avoid some intermediate allocations

[codecov]

Codecov Report

❌ Patch coverage is 0% with 19 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/factorizations/truncation.jl	0.00%	19 Missing ⚠️

Files with missing lines	Coverage Δ
src/factorizations/truncation.jl	17.04% <0.00%> (-70.32%)	⬇️

... and 29 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Jutho]

I think that implementation is very clean. Not sure I can explain all the errors. Some seem to originate from the eigenvectors of a DiagonalTensorMap also being diagonal, which is being changed, right? But other errors I cannot directly explain without running the code locally.

[lkdvos]

I will have a look later, I have a local branch to fix the diagonal implementations already, we'll see what remains after.
I'll also try if it actually works on GPU, and get a similar implementation going for the truncerror version.