Update ensemble interface for SciMLBase v3

[ChrisRackauckas-Claude]

Summary

• Updates DiffEqGPU's ensemble solve implementation to match SciMLBase v3's new ensemble interface
• SciMLBase v3 (via SciML/SciMLBase.jl#1252) redesigned the ensemble RNG system with breaking signature changes
• Updates SciMLBase compat bound from 2.144 to 3

Breaking Changes Addressed

SciMLBase v3 changed:

	v2 (old)	v3 (new)
prob_func	(prob, i, repeat)	(prob, ctx::EnsembleContext)
output_func	(sol, i)	(sol, ctx::EnsembleContext)
solve_batch	(prob, alg, ensemblealg, II, pmap_batch_size; kwargs...)	(prob, alg, ensemblealg, II, pmap_batch_size, ensemble_rng_state; kwargs...)

Where EnsembleContext bundles sim_id, repeat, worker_id, sim_seed, rng, and master_rng.

Changes

src/solve.jl:

• __solve: accepts seed, rng, rng_func kwargs; generates per-trajectory seeds via SciMLBase.generate_sim_seeds; bundles ensemble_rng_state for CPU offload path
• batch_solve: creates EnsembleContext via new _make_ensemble_context helper for each trajectory; calls prob_func(prob, ctx) and output_func(sol, ctx) with context instead of raw indices
• solve_batch (EnsembleThreads override): accepts and forwards ensemble_rng_state positional arg to match SciMLBase v3's solve_batch signature

src/algorithms.jl: Updated docstring examples

test/: Updated all 17 test files to use new (prob, ctx) / (sol, ctx) signatures, replacing i with ctx.sim_id where the trajectory index was used

Project.toml: SciMLBase = "3"

Note

This PR supersedes #431 (dependabot compat-only bump). A compat bump alone is insufficient because DiffEqGPU implements its own __solve/batch_solve that bypass SciMLBase's default batch_func, so the new EnsembleContext creation and RNG infrastructure must be replicated here.

CI will not resolve until downstream deps (e.g., LinearSolve) also bump their SciMLBase compat to v3.

Test plan

• CI passes once LinearSolve and other deps are updated for SciMLBase v3
• Verify GPU ensemble solves produce correct results with default prob_func
• Verify custom prob_func(prob, ctx) with ctx.sim_id works
• Verify seed kwarg produces reproducible GPU ensemble results
• Verify CPU offload path forwards ensemble_rng_state correctly

🤖 Generated with Claude Code