docs: production-deployment guide for trained FLAML models

[immu4989]

Why are these changes needed?

Adds a new Use-Cases/Production-Deployment.md page covering the train → save → reload → predict on new data lifecycle, with a focus on the gotchas that actually surface in production but aren't in the quick-start tutorials.

Each section is grounded in a real user-reported pain point — the page exists because the same handful of issues keep getting filed against FLAML at the inference boundary:

Issue	Pain	Page section
#1101	Categorical encoding silently drifts between fit and predict	§3 — Categorical features at inference time
#1136	automl.model.estimators_[i].predict(raw_X) fails	§4 — Ensemble component access
#1115	MultiOutputRegressor(AutoML()) ignores X_val	§6 — Multi-output regression
#1181	How to persist a StackingRegressor ensemble	§1 — Save and reload
#887	sample_weight + split_type="time" AttributeError	§5 — Sample weights
#1200	SMOTE with cross-validation	§5 — note on imbalanced classification
#1540	Reproducibility audit follow-up	§7 — Versioning and reproducibility

The page follows the existing Use-Cases/ style (it sits next to Task-Oriented-AutoML.md, Zero-Shot-AutoML.md, and Tune-User-Defined-Function.md) and is picked up automatically by the sidebar ({type: 'autogenerated', dirName: 'Use-Cases'} in website/sidebars.js).

Pre-flight verification

Every runnable snippet on the page was exercised against current main before writing. One discovery from that pre-flight is worth flagging in review: the MLflow autolog example in Best-Practices.md, as written, reloads as an unfitted Pipeline on recent MLflow versions (verified on mlflow==2.22.1). The new page recommends the explicit mlflow.sklearn.log_model(automl, artifact_path="...") pattern instead, which round-trips correctly. Happy to file a follow-up bug for the autolog reload path if useful.

What the page does not cover

• Training-time configuration (covered in Task-Oriented-AutoML.md).
• Zero-shot estimators (covered in Zero-Shot-AutoML.md).
• Distributed / Spark / Microsoft Fabric deployment.

Related issue / PR list

This page references but does not duplicate:

• PRs in the reproducibility audit series (#1541, #1546, #1547, #1549, #1551, #1552, #1556)
• The preprocessing-API exposure (#1497)
• Recent inference-path fixes (#1554, #1558, #1561)

Checks

• I've used pre-commit to lint the changes in this PR (note the same in integrated in our CI checks).
• I've included any doc changes needed for https://microsoft.github.io/FLAML/. See https://microsoft.github.io/FLAML/docs/Contribute#documentation to build and test documentation locally.
• I've added tests (if relevant) corresponding to the changes introduced in this PR.
• I've made sure all auto checks have passed.