Adapt to AMReX 26.02 Array4 API changes

[lwJi]

AMReX 26.02 refactored Array4 into a specialization of the new ArrayND class (AMReX-Codes/amrex#4900). This changed the public member API that CarpetX relies on:

• the Dim3 begin/end members became IntVectNDs, so vars.begin.x etc. no longer compile;
• the public jstride/kstride/nstride members were replaced by a get_stride<d>() accessor;
• the new headers use C++20 requires clauses, so building against AMReX ≥ 26.02 needs -std=c++20.

Changes

• CarpetX: use the amrex::lbound/amrex::ubound free functions instead of the begin/end members in interpolate.cxx, io_tsv.cxx, and reduction.cxx. These functions exist in all AMReX versions, so this compiles against both old and new AMReX. Note that ubound is inclusive while the old end member was exclusive; loop bounds and asserts are adjusted accordingly.
• ODESolvers: use Array4::get_stride() when building against AMReX ≥ 26.02, guarded by AMREX_RELEASE_NUMBER (the same pattern as the existing guard in CarpetX/src/fillpatch.cxx); the old stride members are kept for older AMReX.
• scripts: bump the CI configs from C++17 to C++20, required for AMReX ≥ 26.02 and harmless for older versions.

No behavior change is intended; this is a pure API adaptation.

Tests

Built and ran the CarpetX testsuite against AMReX 26.07 (g++, macOS), which exercises the new (AMREX_RELEASE_NUMBER >= 260200) code path. The pre-26.02 code path in ODESolvers is textually unchanged and continues to compile against the AMReX shipped in the current CI containers.

These changes were originally developed and CI-tested (CPU/CUDA/ROCm/oneAPI, real32/real64) in lwJi#103.