Adds sklearn 1.9.0 DLC image (AL2023 + Python 3.12), mirroring xgboost 3.2.

What's added

• docker/sklearn/1.9-0/ — Dockerfile, pyproject.toml, uv.lock. Java 8 Corretto for MMS, which for sagemaker-training, protobuf==3.20.2 pin (required by sagemaker-containers record_pb2).
• .github/config/image/sklearn/sagemaker-1.9-0.yml + .github/workflows/sklearn.pr-1.9-0.yml — image config + PR trigger.
• test/security/data/ecr_scan_allowlist/sklearn/sklearn-1.9.0.json — allowlist (for the protobuf pin).
• test/sklearn/scripts/regen_mme_fixtures.py — seeded regen script for MME S3 fixtures.

Test-fixture hygiene

• S3 MME fixtures moved under mme_models/<sklearn_version>/mme_model_{0,1}.tar.gz so 1.4-2 and 1.9 don't share pickled models. 1.4-2 keeps a byte-copy of the existing fixtures; 1.9 has freshly regenerated ones.
• Same treatment for code/pandas-parquet-file-<slug>.tar.gz.
• test_inference_mme.py numerically asserts the endpoint response against a per-version BASELINE_PREDICTIONS map (via math.isclose(rel_tol=1e-6)); versions without a baseline keep the existing smoke assertion.
• check_versions.py Now accepts pyproject.toml in addition to requirements.txt, and enforces [project].requires-python. Workflow falls back to pyproject when no requirements.txt is present.

Test plan

• 1.4-2-py312 pipeline: local-integ + sagemaker (all 5) → green. (build/unit/sanity/security correctly skipped — no changes under docker/sklearn/1.4-2-py312/.)
• 1.9.0 pipeline: build + unit + local-integ + sanity + security + sagemaker (all 5) → green.
• MME on 1.9: endpoint bytes match BASELINE_PREDICTIONS["1.9.0"].
• MME on 1.4-2: smoke assertion holds via versioned prefix.

Tests Passed

3414b7e