Add TabICL-v2 solver and UEA multivariate dataset loader

[ievred]

Adds a TabICL-v2 solver (tabicl.py) that uses the TabICL tabular foundation model for in-context learning classification on UCR datasets.

• UCR samples (T, 1) are flattened to (N, T) tabular rows and fed directly to TabICLClassifier
• Model loading (and checkpoint download) happens in set_objective — excluded from benchmark timing
• sampling_strategy = "run_once" — classification is solved in a single ICL forward pass
• CUDA/CPU device auto-detected via torch.cuda.is_available(), matching Mantis
• Validated on ECG5000: accuracy 94.6 %, f1_weighted 0.940 in 74 s on CPU
• 4 unit tests added in test_tabicl.py

[ievred]

UPD:

Changes

tabicl.py — TabICL-v2 solver for UCR/UEA classification

• Uses tabicl (ICML 2026), a tabular in-context learning foundation model
• UCR/UEA samples (T, C) are flattened to (N, T×C) tabular rows — no interpolation needed for fixed-length archives
• Model instantiation and checkpoint download happen in set_objective (excluded from timing); fit runs in run
• sampling_strategy = "run_once" — classification is a single ICL forward pass
• CUDA/CPU device auto-detected via torch.cuda.is_available(), matching the Mantis solver pattern
• Checkpoint reload guard prevents re-downloading across consecutive datasets

uea.py — UEA multivariate dataset loader

• Mirrors ucr.py exactly, backed by the same tslearn.UCR_UEA_datasets loader
• Default dataset: BasicMotions (T=100, C=6, 4 classes)
• Covers all 30 multivariate UEA datasets available via tslearn

test_tabicl.py — 4 unit tests (offline, fake tabicl)

• fit shapes, predict shapes, classifier reuse across datasets, skip on unsupported tasks

Results

Dataset	Accuracy	Balanced Acc	F1 weighted	Time
ECG5000 (UCR, univariate)	0.946	0.601	0.941	74 s (CPU)
BasicMotions (UEA, multivariate)	1.000	1.000	1.000	2 s (cached)