refactor(tent): canonicalize AMD GPU location prefix to hip (#3532)

- Rename AMD GPU location/device prefix from "rocm" to "hip" to match classic TE (GPU_PREFIX)
- Keep "rocm" as a parse-only alias so existing TENT configs and custom NIC matrices still work
- Add shared helpers (isAmdGpuLocationType, makeAmdGpuLocation, memTypeFromLocation) and use them across allocators, memory prober, transport selector, and plugin dispatch
- Move the AMD prefix constants and isAmdGpuLocationType into a new self-contained header tent/runtime/amd_location.h so the out-of-tree rocm device plugin shares the same single source of truth instead of duplicating them
- Canonicalize legacy "rocm:N" matrix keys to "hip:N" at the topology storage/query boundary (canonicalizeLocation) so per-GPU NIC priorities configured under the legacy key are not silently ignored; specifying both "rocm:N" and "hip:N" for the same device is rejected as a conflict
- Update docs and tests to reflect hip:N output and rocm:N compatibility

Signed-off-by: staryxchen <staryxchen@tencent.com>