[Build] Find ROCm through rocm-sdk before falling back to /opt/rocm*

[xinyazhang]

Motivation

Since ROCm 7.14, ROCm can be installed through pip. The HIP package used to build
FlyDSL must come from the selected ROCm installation. Passing a ROCm prefix in
find_package(hip ... PATHS ...) did not guarantee this: CMake searches its
normal package locations before PATHS, so a system HIP could win over the SDK
in the active virtual environment.

Change

The ROCm runtime now tries HIP in this order:

1. ROCM_PATH, if it resolves a HIP package;
2. the root reported by rocm-sdk path --root, if that command succeeds and
   resolves a HIP package;
3. CMake's normal package locations and the existing /opt/rocm* fallback.

The first two searches use CMake's NO_DEFAULT_PATH option so defaults cannot
precede the selected prefix. An empty or stale ROCM_PATH falls through to the
SDK. While loading HIP from the SDK, ROCM_PATH temporarily points to that SDK
so HIP's own dependencies are found there too; its original value is restored
afterwards. If rocm-sdk is missing or fails, the original fallback still works.

As with other CMake packages, an explicit or previously cached hip_DIR remains
authoritative. Clear hip_DIR or use a fresh build directory when switching
ROCm installations.

Validation

• The original PR built flydsl-0.3.1.dev1129-cp313-cp313-linux_x86_64.whl
  in a Debian 13 container with a ROCm 7.14 virtual environment.
• Configure checks with the installed pip ROCm SDK and a fake system HIP
  confirm that unset, empty, and stale ROCM_PATH values select the SDK.
  A valid ROCM_PATH wins, and a failed rocm-sdk command falls back to
  normal CMake package discovery.
• The existing CMake backend tests pass (4 tests).
• After updating from main, a local Sphinx build with warnings treated as
  errors passed. This includes the docs change that removes the remote mapping
  request which had hit a GitHub API rate limit on the older branch.
• CI on the earlier search-order fix built a PR wheel; MI325, MI35x, and Navi
  GPU tests and benchmarks passed. Navi passed on retry after its runner failed
  to download the wheel artifact on the first attempt. Checks for the current
  head are shown below.

Submission Checklist

• Look over the contributing guidelines at
  https://github.com/ROCm/TheRock/blob/main/GOVERNANCE.md#pull-requests.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[coderfeli]

Reviewed the updated HIP discovery change. The real pip ROCm HIP config reads ROCM_PATH when resolving dependencies, so an empty or stale value previously caused configure to fail at AMDDeviceLibs even after hip_DIR selected the SDK. Commit f39f41b scopes ROCM_PATH to the SDK during that lookup and restores it afterward. I also removed the mock-only discovery test.

I configured against the installed pip ROCm SDK with ROCM_PATH unset, empty, and stale; all three now select the SDK. A valid explicit prefix and the normal fallback after a failed rocm-sdk command also resolved as expected. The four existing CMake backend tests pass. I found no remaining blocking issue in this diff; CI for the updated head is still pending.