Improve latentneuralode

[robin-janssen]

Improve LatentNeuralODE performance:

• Use float64 only for the numerical solver, reducing the model's memory footprint and latency of forward and backward pass
• Optionally, LNODE now also supports the torchode PIDcontroller and adjoint (or joint adjoint) method for the backward pass (which is much slower but also very memory-friendly, cf. the torchode docs)
• The computation of the auxiliary losses was improved: First and second derivative losses now use proper central differences formulae.

Additionally, this branch contains a first implementation of the iterative evaluation, aimed at investigating error propagation across iterative calls to the solver. This will be further refined in branch improve-surrogate-evaluation.

[codecov]

Codecov Report

❌ Patch coverage is 71.27660% with 81 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...es/surrogates/LatentNeuralODE/latent_neural_ode.py	54.23%	54 Missing ⚠️
...surrogates/AbstractSurrogate/abstract_surrogate.py	30.00%	14 Missing ⚠️
codes/tune/optuna_fcts.py	74.19%	8 Missing ⚠️
codes/benchmark/bench_fcts.py	95.58%	3 Missing ⚠️
codes/surrogates/LatentPolynomial/latent_poly.py	97.22%	1 Missing ⚠️
codes/train/train_fcts.py	0.00%	1 Missing ⚠️

Files with missing lines	Coverage Δ
codes/surrogates/DeepONet/deeponet.py	90.22% <100.00%> (+0.07%)	⬆️
codes/surrogates/FCNN/fcnn.py	92.24% <100.00%> (+0.34%)	⬆️
codes/utils/data_utils.py	58.04% <ø> (ø)
codes/surrogates/LatentPolynomial/latent_poly.py	77.77% <97.22%> (+2.91%)	⬆️
codes/train/train_fcts.py	66.90% <0.00%> (ø)
codes/benchmark/bench_fcts.py	83.69% <95.58%> (+1.48%)	⬆️
codes/tune/optuna_fcts.py	78.65% <74.19%> (-0.72%)	⬇️
...surrogates/AbstractSurrogate/abstract_surrogate.py	61.78% <30.00%> (-0.88%)	⬇️
...es/surrogates/LatentNeuralODE/latent_neural_ode.py	66.24% <54.23%> (-10.36%)	⬇️

🚀 New features to boost your workflow:

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