docs: add custom forecaster tutorial (#1128)

[charanteja25]

Closes #1128

Is your feature request related to a problem?

Yes, issue #1128. There is no tutorial that shows how to write a custom Forecaster. The only custom example, examples/benchmarks/custom/custom_forecaster.py, is a template for the BEAM benchmark interface (BacktestForecasterMixin). A user who wants to add their own model to a ForecastingModel has to read the source code of the built-in forecasters.

Describe the solution you'd like

A new tutorial, examples/tutorials/custom_forecaster.py (with the paired .ipynb), that shows:

• The contract of a Forecaster: hparams, is_fitted, fit, predict, and the quantiles and horizons fields.
• A small working example: a weekly profile forecaster. For each time slot of the week, it stores the quantiles of the past values.
• How to use it in a ForecastingModel and a CustomForecastingWorkflow.
• A comparison with the built-in ConstantQuantileForecaster.

The tutorial is linked in docs/source/examples.rst (table and toctree), after "Custom Pipeline".

Describe alternatives you've considered

• Extend the BEAM template in examples/benchmarks/custom: it is about benchmarking, not about the model interface. The tutorial links to it.
• Write a how-to guide in docs/source/user_guide/guides/: a tutorial with runnable code fits the other tutorials better. A short link from a guide can follow.
• Use a gradient boosting model as the example: it needs feature engineering, which hides the contract. The weekly profile uses only the target column.

Affected package(s)

• docs
• examples / deployment
• openstef-core
• openstef-models
• openstef-beam
• openstef-foundation-models
• openstef-meta

Would this require breaking changes to the public API?

No. Only a new tutorial and one docs index entry.

Additional context

I read the issue as a request for a tutorial on the Forecaster interface (a custom model). If you meant something else, please tell me and I change it.

Result on the Liander data used in the other tutorials (45 days of training, 36 h forecast):

Model	MAE of the median forecast
ConstantQuantileForecaster	229,816
WeeklyProfileForecaster	183,425

This is one 36 h forecast window. It is an illustration for the tutorial, not a benchmark.

Notes for the review:

• ForecastingModel holds back validation and test data by default. The weekly profile needs every week, so the tutorial sets DataSplitter(val_fraction=0.0, test_fraction=0.0) and explains why in a note.
• I removed the unit from the plot axis ("Load"). The values reach ±1,000,000 and can be negative, so "MW" looked wrong. Please correct me if the unit is known.

Checks:

• poe all --check passes (ruff, ty, tests, doctests, REUSE, notebooks).
• On my machine (timezone UTC+05:30), 5 existing tests fail in test_datetime.py and test_evaluation_pipeline.py. They pass with TZ=UTC. They are not related to this change.
• Commit is signed off (DCO).