Unify spline eval via many-point kernel and keep legacy APIs as harness

[krystophny]

Goal

Make the new GPU-ready many-point batch spline evaluation the single source of truth, and route existing scalar + batch entry points through it without changing behavior.

Motivation

• Reduce duplication across scalar and batch spline evaluation code paths.
• Ensure OpenACC acceleration can be used by downstream codes through a consistent interface.
• Keep the old routines as a behavioral harness (tests) to guarantee no regressions.

Prerequisite

• Fix scalar periodic wrapping inconsistency (issue Scalar periodic spline eval ignores x_min #201) so scalar and batch semantics are aligned before unification.

Proposed work

1. Introduce an internal evaluation kernel (per dimension) that supports:
   • single point (npts=1) and many points
   • output layout y(nq, npts)
2. Implement facades:
   • existing scalar APIs call the kernel with nq=1/appropriate layout
   • existing batch single-point APIs call the kernel with npts=1
   • existing many-point APIs already use the kernel
3. Keep the old public routines as thin wrappers initially, but add tests that cross-check:
   • old scalar vs new facade
   • old batch single-point vs new facade
   • old many-point vs reference
4. Ensure OpenACC pass-through:
   • keep resident variants that rely on present(...)
   • document the downstream-managed data lifetime requirements

Acceptance

• No numerical behavior change (existing tests remain unchanged; new cross-check tests added).
• Downstream can compile+link with OpenACC and keep data resident across algorithms.

[krystophny]

Duplicate of #204.