[NDTensors] Some fixes for element type promotion and conversion

[mtfishman]

This fixes a few issues with element type promotion and conversion, such as mixed element type contractions of block sparse tensors on GPU as well as better element type preservation when the noise term is enabled in DMRG.

I also extended the new DMRG tests on GPU introduced in #1236 to the case of QN conservation/block sparse tensors, which pass locally for me using a Metal backend.

• Add a test for an eigendecomposition of a complex Hermitian matrix, testing that the eigenvalues are real.

[codecov-commenter]

Codecov Report

Attention: 3 lines in your changes are missing coverage. Please review.

Comparison is base (b7119d6) 85.44% compared to head (c5c8175) 54.81%.
Report is 1 commits behind head on main.

❗ Current head c5c8175 differs from pull request most recent head 5a20441. Consider uploading reports for the commit 5a20441 to get more accurate results

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #1244       +/-   ##
===========================================
- Coverage   85.44%   54.81%   -30.63%     
===========================================
  Files          89       88        -1     
  Lines        8400     8348       -52     
===========================================
- Hits         7177     4576     -2601     
- Misses       1223     3772     +2549     

Files	Coverage Δ
src/mps/dmrg.jl	0.00% <0.00%> (-84.04%)	⬇️

... and 36 files with indirect coverage changes

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[mtfishman]

@kmp5VT so interestingly the new tests I added for block sparse DMRG pass on CPU and Metal (I checked locally) but seem to be failing with CUDA.

[kmp5VT]

@kmp5VT so interestingly the new tests I added for block sparse DMRG pass on CPU and Metal (I checked locally) but seem to be failing with CUDA.

That is interesting, I can look into what is going on later today!

[mtfishman]

Thanks! I've marked the tests as broken for now so I can merge this fix.

[mtfishman]

@kmp5VT I noticed some of the tests you wrote in #1236 used arrays with uninitialized memory (i.e. Vector{Float64}(undef, 4)) and then used those in algebra operations. In general that is not a good idea since then you can randomly get elements with NaN or Inf which don't play nicely when doing algebra. Better to allocate them to something random, like randn(Float64, 4).