🚀 Feature

Implement time series forecasting evaluation metrics (e.g. SoftDTW)

Motivation

Current time series evaluation relies primarily on point-wise metrics (MAE, RMSE) which don't capture temporal alignment quality. Dynamic Time Warping (DTW) based metrics better assess forecast shape accuracy but aren't differentiable. SoftDTW (Cuturi & Blondel 2017) provides a differentiable alternative - implementing these would enable better optimization and evaluation for neural forecasting models.

Pitch

1. Add SoftDTW and related time series metrics (soft_dtw_loss, etc)
2. Include gradient-computation support for training
3. Add benchmark comparison against traditional metrics
4. Document usage examples for both evaluation and training

Alternatives

• tslearn, which seems to be unmaintained.

Additional context

N/A