#1538 tests: skip expert_cache_segmented_test on HIP builds (#1475)

本文

## Summary

- `expert_cache_segmented_test` is registered with `SKIP_RETURN_CODE 77` and skips cleanly on missing devices or missing driver VMM support — but on HIP builds `open_segmented` is compiled out (`STRATA_EC_NO_VMM`) and refuses with `--vram-elastic (a segmented expert cache) is CUDA-only for now`, which fell through to `fail("open")` and showed as a ctest failure on every AMD run.
- Match the CUDA-only refusal alongside the VMM one so the test reports skip (77) like the other backend-specific tests.

Fixes #1475.

## Test plan

- [x] `g++ -std=c++20 -fsyntax-only` on the test via the HIP compat include path: clean
- [ ] HIP ctest run: `expert_cache_segmented_test` shows "Skipped" instead of "Failed" (needs an AMD machine — reporter's R9700 environment per #1475)
- [ ] CUDA build: unchanged — the wait/`cudaGetDeviceCount` and VMM paths behave exactly as before