#1828 sycl: register iq_parity and ple_parity like the top-level build

Descripción

Fixes #1795.

## What

The port's generic parity loop registered every test as `<test> --selftest`. For `iq_parity` that string became its fixture-directory argument (`argv[1]`), and for `ple_parity` the GGUF/capture-bin relative paths resolved against the build tree — so both failed on every checkout for reasons unrelated to the kernels, exactly as the reporter showed.

This takes `iq` and `ple` out of the `foreach` and gives them the same registrations the top-level `CMakeLists.txt` uses:

- `iq_parity`: the `tools/iq_fixture.py` CTest fixture (generated into the build tree), the same python-with-numpy interpreter probe (`IQ_FIXTURE_PYTHON` overrides it), `FIXTURES_SETUP`/`FIXTURES_REQUIRED`, and exit-3 as a skip. The test itself runs as `iq_parity <build>/iq_fixture`.
- `ple_parity`: `WORKING_DIRECTORY` = the engine root so the binary's documented relative defaults (`pack/full`, `bench/micro`, `../../Q2_0/...`) resolve; `--in`/`--out` from a `STRATA_PLE_FIXTURE_DIR` cache path (default `bench/micro`), plus new `STRATA_PLE_GGUF` and `STRATA_PLE_PACK` configure-time inputs for the artifacts the reporter asked to name.

Missing inputs stay a hard error inside the test, as the top-level build intends — this only makes the error satisfiable, and `docs/INTEL.md`'s parity table now names the variables.

## Checked

- `ctest -N` on the exact parity block (stubbed `strata_resolve`/targets): `iq_parity_fixtures` + `iq_parity` and `ple_parity` each register exactly once; with `-DSTRATA_PLE_GGUF`/`STRATA_PLE_PACK` the command line gains `--gguf`/`--pack`, without them it is clean of stray arguments; the fixtures command is `<python> <root>/tools/iq_fixture.py --out <build>/iq_fixture`.
- Not run: a real `ctest` — this machine has no SYCL toolchain; the registration logic is what changed, and it is verified to mirror the top-level contract. A port user can confirm with the issue's repro (`ctest -R '^(iq|ple)_parity$'`).