[gpu-2] Benchmark script for VlasovAmpere and VlasovMaxwell

[max-models]

Solves the following issue(s):

The "first CPU vs GPU timing" item of #689, tracked in #650. This PR provides the script and a CPU baseline. The GPU numbers come from running it on the H100 (#687).

Stack: gpu-2. Merge it after gpu-1 (gpu-1-vlasov-ampere-e2e), which is its base, once all the CUDA PRs are in.

Core changes:

New script profiling/benchmarks/vlasov_benchmark.py. It runs a fixed number of time steps of a standard problem with VlasovAmpereOneSpecies or VlasovMaxwellOneSpecies, on the NumPy or the CuPy backend, and reports:

• setup time, CUDA compile time, time per step (median and first step), diagnostics and output time;
• time per propagator, and ns per marker and step;
• host/device transfers in the time loop (cunumpy.profiling.count_transfers);
• optionally all of it as JSON (--json), to compare runs.

Cases:

case	models	setup
landau	both	weak Landau damping as in the verification tests: box of length 4π (k = 0.5), 1e-3 density perturbation, control-variate weights, cubic splines in x
weibel	VlasovMaxwell	Weibel instability as in Kraus et al., J. Plasma Phys. 83 (2017), Sec. 6.2: box 2π/1.25, vth1 = vth3 = 0.02/√2, vth2 = √12 vth1, seed B3 = -1e-4 cos(kx)

All runs use alpha = 1, epsilon = -1, as the verification tests do. The initial magnetic energy of the Weibel seed matches the analytic value (1.2551e-8 vs 1.2566e-8). The growth rate was not checked, since that needs thousands of steps.

Sizes: smoke (8 cells × 50 ppc, 3 steps), standard (32 × 1000, 20 steps), large (128 × 20000 = 2.56M markers, 20 steps). Grid, degree, ppc, steps and dt can all be overridden.

Timing details:

• Every timed region ends with xp.synchronize(), and times are the maximum over MPI ranks.
• Per-propagator timing adds one synchronization per propagator and step; turn it off with --no-propagator-timing.
• Sorting boxes are off by default, because of SortingBoxes cannot be created on the CuPy backend #726.
• On the fake CuPy (CUNUMPY_FAKE_CUPY=1 ... --backend cupy --size smoke), the CUDA launches are emulated on the CPU. That checks the device code path without a GPU, but its times are not GPU timings, and the report says so.

CPU baseline (NumPy, --size standard: 32k markers, 20 steps, 1 rank, Apple M-series laptop):

run	ms per step	ns per marker and step	VlasovAmpereCoupling share
VlasovAmpere, landau	67	2087	86 %
VlasovMaxwell, landau	82	2571	68 %
VlasovMaxwell, weibel	86	2678	68 %

Testing: all three cases at --size smoke and --size standard on NumPy; VlasovMaxwell weibel at --size smoke on the fake CuPy (all four propagators ran, with no host kernels and no host-to-device copies in the time loop). Not run on a GPU.

Documentation changes: the script's docstring (usage, cases, sizes, CPU reference).

🤖 Generated with Claude Code

[spossann]

Let's use the scope-profiler instead.