[32-bit] nvfp4_plan_cache.h fails with size_t narrowing on i586

Problem

The v0.0.2 source does not compile for i586/32-bit x86 with GCC and -Werror=narrowing. The failure is in src/vt/cuda/nvfp4_plan_cache.h:

constexpr size_t kGolden = sizeof(size_t) == 8 ? size_t{0x9e3779b97f4a7c15ULL}
                                              : size_t{0x9e3779b9UL};

GCC reports:

error: narrowing conversion of ‘11400714819323198485’ from ‘long long unsigned int’ to ‘size_t’ {aka ‘unsigned int’} [-Wnarrowing]

The conditional expression is valid conceptually, but the 64-bit braced initializer is still diagnosed as narrowing when size_t is 32-bit.

Reproducer

Build v0.0.2 on i586 with the normal openSUSE optimization flags and C++20 enabled.

Proposed direction

Use explicit casts or architecture-conditional definitions so both constants are well-formed for 32-bit and 64-bit size_t, for example:

constexpr size_t kGolden = sizeof(size_t) == 8
    ? static_cast<size_t>(0x9e3779b97f4a7c15ULL)
    : static_cast<size_t>(0x9e3779b9UL);

The openSUSE package currently carries this as a downstream patch.

Start with src/vt/cuda/nvfp4_plan_cache.h and inspect the kGolden definition. Build v0.0.2 for i586 with the normal openSUSE optimization flags, C++20, and -Werror=narrowing to reproduce the failure. Done means the header compiles on both 32-bit and 64-bit size_t targets without narrowing diagnostics.