Perf Optimize spectral operations and backend dispatch

[lucas-heck]

This PR reduces intermediate allocations, mainly by using PyTorch optimized operations, and backend dispatch overhead in a couple of layers:

• Reconstruct matrices by scaling eigenvector columns instead of building diagonal tensors, preserving the CUDA intermediate layout.
• Keep Loewner thresholds on-device and use dense masking on accelerators while retaining indexed updates on CPU.
• Cache LogEig/ExpEig backend selection and restore it when loading checkpoints.

It also skips matrix-log clamping warnings for CUDA inputs requiring gradient, which might be a maintainer decision but this way it avoids synchronization during training (I don't mind changing back).

Validation:

• Numerical tests and gradchecks passed on CPU and CUDA
• Standalone GPU training benchmarks reduced median fit time by 4.6%/4.7% for SPDNet and 4.5%/5.6% for EEGSPDNet on two machines, respectively, while CPU training time was pretty much the same.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 64.70588% with 12 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
spd_learn/functional/autograd.py	56.25%	4 Missing and 3 partials ⚠️
spd_learn/modules/modeig.py	62.50%	1 Missing and 2 partials ⚠️
spd_learn/functional/core.py	75.00%	1 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!

[bruAristimunha]

Small details, and all good

[bruAristimunha]

Thanks a lot @lucas-heck, really nice work, and the benchmark numbers in the description help a lot.

I did a second full pass before merging:

• Full test suite on all four open PRs merged together (#42 → #43 → #44 → #45): 825 passed, 147 skipped, plus the docs code blocks.
• The non-CPU torch.where branch in modeig_backward only runs on accelerators, so CI does not cover it (that explains most of the codecov gap). I ran it on Apple MPS with a batch that includes repeated eigenvalues: matrix_log gradients match CPU to 6e-8.
• LogEig / ExpEig: after load_state_dict from a checkpoint with a different autograd flag, the cached backend is correctly picked up.
• gradcheck on batched float64 matrix_log passes.

Small, non-blocking notes:

1. get_loewner_threshold conflicts with #45, which makes the threshold per matrix. I'll keep #45's version, since it is a superset of this one (tensor, on-device, per matrix).
2. The CHANGELOG.md diff also reformats the old model/metric tables (whitespace only). That's harmless, just noise in the diff.
3. The _use_autograd cache only updates on __init__ and load_state_dict. If someone changes autograd_ in place, it goes stale. That's fine as long as the buffer is treated as read-only.

Merging. Thanks again!