Context

TimeSeriesDataSet supports multi-target forecasting, but the current LSTMModel example (and implementation) is effectively univariate, which can lead to confusion and runtime shape mismatches when multiple targets are provided.

An older PR (#1449) attempted to address this, but it is now stale relative to main.

Goal

Clarify and/or extend LSTMModel support such that behavior is explicit and consistent:

• Either clearly documented as univariate only, or
• Provide native multi-target (multivariate) support

Proposed Approach

1. Review [ENH] Support tuning any model and extend LSTMModel in docs to support multi-target datasets #1449 against current main to identify:
   • What logic is still applicable
   • What no longer applies due to recent refactors
2. Define a minimal, backward-compatible API for multivariate targets, including:
   • Decoder output dimensionality
   • Loss handling
   • Target normalization
3. Decide whether documentation clarification should be done in parallel

Questions for Maintainers

• Is native multivariate support for LSTMModel desired in principle?
• If yes, should this be a single model with n_targets > 1 or separate heads?
• Should documentation clarification be merged independently?

Happy to proceed with implementation once aligned.