Avoid CUDA Host/Device Constexpr Warning in Vector Reduction Identities

Problem

When building IPPL with CUDA on GH200, BareField tests emit warning #20013-D from the Kokkos::reduction_identity<ippl::Vector<T, Dim>> specialization in src/Field/BareField.hpp.

The warning is triggered by calling std::numeric_limits<T>::infinity() inside KOKKOS_FORCEINLINE_FUNCTION methods:

KOKKOS_FORCEINLINE_FUNCTION static ippl::Vector<T, Dim> min() {
    return ippl::Vector<T, Dim>(std::numeric_limits<T>::infinity());
}

KOKKOS_FORCEINLINE_FUNCTION static ippl::Vector<T, Dim> max() {
    return ippl::Vector<T, Dim>(-std::numeric_limits<T>::infinity());
}

NVCC diagnoses this because std::numeric_limits<T>::infinity() is treated as a constexpr host function being called from a host-device function.

We should avoid requiring the compiler workaround:

--expt-relaxed-constexpr

Proposed Fix

Delegate the scalar identity values back to Kokkos:

KOKKOS_FORCEINLINE_FUNCTION static ippl::Vector<T, Dim> min() {
    return ippl::Vector<T, Dim>(Kokkos::reduction_identity<T>::min());
}

KOKKOS_FORCEINLINE_FUNCTION static ippl::Vector<T, Dim> max() {
    return ippl::Vector<T, Dim>(Kokkos::reduction_identity<T>::max());
}

This matches Kokkos reduction semantics and keeps vector field reductions consistent with scalar field reductions. For float and double, Kokkos uses finite extrema (FLT_MAX, DBL_MAX) rather than infinities, which avoids the CUDA host/device constexpr warning.