[ShawnL00]

This implementation adapts the GPU scheduling ideas from Isuru’s L2P optimization in #158:

• One workgroup per target box.
• Shared local expansion coefficients in local memory.
• Diff: configurable workgroup size.; uses each box’s target count instead of requiring max_ntargets_in_one_box.

(Assisted by ChatGPT and Claude.)

To-do: add recurrences for mixed derivatives.