Release validation logs (T7)

One JSON file per released version and GPU, written by

$MMAEMU_GPU_LOCK timeout 600 python tools/validate_release.py dist/mmaemu-<version>-py3-none-any.whl

and committed as part of the release (docs/release.md, section 3). File name: <version>_<gpu>.json, where <gpu> is the device name without vendor words plus its capability (for example rtx4060-sm89).

Format mmaemu-release-validation/1:

Key	Content
wheel	file name, sha256, size, version, the fatbin modules it carries
host, source_tree_commit, created_utc	where and from which commit the validation ran
venv	how the fresh environment was built (uv or venv + pip) and which torch was installed
device	GPU name, capability, SM count, torch and CUDA versions inside the fresh environment
checks	environment, venv, install, info, selftest, loader, conformance, evo2: each with status (PASS / FAIL / SKIP), a summary and its evidence
info_before_selftest, selftest, conformance, evo2	the raw reports of python -m mmaemu info, python -m mmaemu selftest --force and the two in-venv checks (conformance.status_after: the backend status after the kernels ran, as auto-selection sees it, so a gated kernel not validated on the device shows source: null even when it ran on explicit request; conformance.modules: the image, source and sha256, that each kernel which ran was loaded from, which the loader check compares with the wheel's manifest)
verdict	PASS only if every check is PASS

Only reports of release candidates and releases are committed here; dry runs go elsewhere (--out).