Fix env image build failures and liveness probe timeouts in share-models notebook CI

[Copilot]

CI was failing in two distinct ways: the SKLearnEnv Docker build was breaking due to EOL Python 3.8 and overly restrictive legacy package pins, and the managed online deployment was hitting liveness probe 502s before the MLflow serving container finished loading.

env_train/Dockerfile

• python:3.8.13 → python:3.10 (3.8 EOL Oct 2024)
• Dropped packages unused by the training script (matplotlib, psutil, tqdm, ipykernel)
• Relaxed version pins to ranges that actually resolve today (pandas>=1.3,<3.0, scipy>=1.7,<2.0, numpy>=1.21,<2.0)

artifacts/model/conda.yaml

• Removed pip<=22.0.4 (breaks modern dependency resolution) and deprecated azureml-ai-monitoring/azureml-contrib-services
• Updated to python=3.10.*, mlflow>=2.0, relaxed remaining pins

Notebook deployment cell

• Replaced the try/except HttpResponseError retry loop with ProbeSettings(initial_delay=300) — the correct fix for a slow-starting MLflow container rather than catching the failure after it's declared dead:

demo_deployment = ManagedOnlineDeployment(
    ...
    liveness_probe=ProbeSettings(failure_threshold=30, timeout=30, period=10, initial_delay=300),
    readiness_probe=ProbeSettings(failure_threshold=10, timeout=30, period=10, initial_delay=300),
)
ml_client_workspace.online_deployments.begin_create_or_update(demo_deployment).result()