GPU adaptive step size

[samtalki]

AdaptiveStepSize currently throws on _GPUStyle() paths (step_rule
dispatch in src/solver.jl). per-problem backtracking does scalar
comparisons that aren't GPU-friendly without rework.

• per-problem backtracking via masked broadcast over the active subset
• device-resident Vector{Bool} mask reused across iterations
• verify the lipschitz-estimate update doesn't require scalar comparison
• device test under MARGUERITE_TEST_GROUP=gpu

cross-vendor: should work uniformly on Metal / CUDA / AMDGPU once the
broadcast pattern is right.