docs: revert getting-started.md tutorial to xgb_train (pytorch_train crashes on Apple Silicon)

[sallycr]

What type of PR is this? (check all applicable)

• Refactor
• Feature
• Bug Fix
• Optimization
• Documentation Update

What changed?

Reverts docs/user-guides/getting-started/getting-started.md's tutorial to its pre-#1749 state, teaching the in-core XGBoost pipeline (python/examples/pipelines/california_housing_xgb/) again instead of michelangelo-examples' pytorch_train (PyTorch Lightning). No other file from #1749 is touched — docs/getting-started/sandbox-setup.md, docs/user-guides/examples/index.md, and docs/user-guides/ml-pipelines/workflow-patterns.md's repoints to michelangelo-examples are unaffected and still correct.

Why?

#1749 switched the tutorial's flagship worked example to pytorch_train. Its documented local-run command (python -m michelangelo_examples.california_housing.pipelines.pytorch_train.pipeline) crashes on any Apple Silicon Mac:

ValueError: You set `strategy=<...RayDDPStrategy...>` but strategies from the DDP
family are not supported on the MPS accelerator. Either explicitly set
`accelerator='cpu'` or change the strategy.

Root-caused and tracked as issue #1867, with a design already written up (fix requires adding a new accelerator field to core michelangelo's LightningTrainerKwargs — not something fixable from docs or from michelangelo-examples alone, since that dataclass has no such field today). Also confirmed RayDeepSpeedStrategy doesn't sidestep this — it's a subclass of Lightning's own DDPStrategy/ParallelStrategy, so it hits the identical MPS rejection.

Until #1867's fix ships, pytorch_train's local-run path is broken for a likely-large fraction of individual contributors' dev machines (any Apple Silicon Mac). The in-core XGBoost example has no PyTorch/MPS dependency at all and runs clean end-to-end (re-verified locally, no manual workarounds). Reverting the tutorial back to it keeps getting-started.md trustworthy in the meantime; the switch to pytorch_train can be reinstated once #1867 is fixed and released.

How did you test it?

• Confirmed no other commit touched getting-started.md between docs: repoint california_housing doc references to michelangelo-examples #1749's merge and this PR's base, so this is a clean, non-lossy revert of just that file.
• Confirmed the restored content's pipeline (california_housing_xgb) still exists in-core (not yet deleted — chore(examples): remove california_housing_xgb pipeline (migrated to michelangelo-examples) #1751, the PR that removes it, is still open/unmerged) and has no PyTorch/Lightning/MPS/DDP references at all.
• Re-ran the restored tutorial's documented local-run command against a fresh origin/main checkout — completes cleanly, no manual workaround.

Potential risks

None beyond a temporary doc/tutorial content change. No code changes.

Breaking Changes

• No breaking changes
• API changes (Go exported symbols or function signatures, Python public functions or classes)
• Proto changes (enum value renumbering, field number changes, field removal, service removal)
• Helm changes (new required values, renamed or removed keys, changed value semantics)
• Config/deployment changes (new required env vars, renamed container args, changed ports or mount paths)

Release notes

N/A

Documentation Changes

Yes — reverts the getting-started tutorial's flagship example from pytorch_train back to the in-core XGBoost pipeline, until #1867 (pytorch_train crashes on Apple Silicon) is fixed.