Add SMT (Supervised Memory Training) estimator

[kashif]

Adds a torch estimator implementing Supervised Memory Training (SMT, Kumar & Isola 2026, https://arxiv.org/abs/2606.06479) as an alternative to DeepAR.

What it does

• A bidirectional Transformer encoder compresses the past (target lags + covariates) into a memory of learned register tokens.
• A causal decoder predicts each future target from the previous future target and future covariates — a predictive-state objective with a DistributionOutput NLL head.
• A recurrent memory cell is trained to reproduce the teacher's one-step memory transitions (m_t, x_{t+1}) -> m_{t+1}, so the RNN is trained without backpropagation through time.
• At inference the recurrent cell is unrolled with O(1) state and sampled autoregressively.

It reuses the DeepAR feature pipeline (lags, time/age features, static embeddings, mean scaling), so SMTEstimator is a drop-in with the same covariate handling and DistributionOutput heads.

Included

• gluonts.torch.model.smt — module, lightning module, estimator
• unit tests in test/torch/model/test_smt.py
• an electricity vs DeepAR comparison notebook in examples/