GPU-accelerate time-dependent propagation (propagate/propagate_step)

[JackWetherell]

Add GPU=False parameter to propagate() and propagate_step() across all methods, mirroring the existing GPU flag pattern used in the interacting solver (hamiltonian/solve).

interacting.py:

• Add _gpu_expm_action(): Arnoldi/Krylov iteration that approximates exp(c·A)@v using GPU sparse matrix-vector products (CuPy CSR) and a small dense Hessenberg expm computed on CPU via scipy.
• propagate_step(GPU=False): when GPU=True, builds the many-body Vptrb operator with cupyx.scipy.sparse, assembles Hp on GPU, then evolves the wavefunction with _gpu_expm_action instead of expm_multiply.
• propagate(GPU=False): passes GPU flag to hamiltonian() so H is constructed on GPU, creates a CuPy sparse identity for the Kronecker product helpers, and prints the GPU device name.

non_interacting.py:

• propagate_step(GPU=False): when GPU=True, transfers the N×N Hamiltonian to CuPy, performs a single eigh() diagonalisation per spin channel, then applies exp(-i dt e) analytically to every orbital (exact for Hermitian H, avoids repeated expm_multiply calls).
• propagate(GPU=False): prints GPU device name and forwards GPU flag to propagate_step.

Wrapper methods (hartree, hartree_fock, lda, hybrid, kssce):

• Add GPU=False parameter and forward it to non_interacting.propagate.

https://claude.ai/code/session_01TiMh7h9Hdmsdtv8S7TrUtT

[JackWetherell]

@DrMattHodgson I will test this change for GPU performance on the time-dependent code, but do the TD unit tests cover things sufficiently that if they pass I assume it has not changed the result? Cheers!

[DrMattHodgson]

@JackWetherell I would think so. The test compares the one-electron wavefunction from the many-body code to the analytic solution at each time step. So, it doesn't test the interaction at all, and it's limited to the harmonic oscillator with a uniform perturbation, but if your GPU acceleration introduces a bug, I think it would affect all systems and, hence, be caught.

[JackWetherell]

@DrMattHodgson Ok great! I will run some benchmarks locally!

[JackWetherell]

@JackWetherell to resolve conflicts.