Add Timestral wrapper and fev-bench results (5M, 28M, 123M, 209M)

[adrienpetralia]

Timestral is a family of time-series forecasting models from Mistral AI (weights, code).

This PR adds:

• models/timestral/model.py and requirements.txt: wrapper around timestral.TimestralForecaster
• benchmarks/fev_bench/results/timestral-{5m,28m,123m,209m}.csv
• 4 entries in benchmarks/fev_bench/models.yaml

Results on the 100 fev-bench tasks:

model	SQL skill	MASE skill
Timestral-5M	46.19	34.54
Timestral-28M	48.42	37.27
Timestral-123M	50.18	39.19
Timestral-209M	50.39	39.49

To reproduce (one size):

uv run python models/evaluate.py -m timestral -b benchmarks/fev_bench/tasks.yaml -n Timestral-209M -k '{"model_path": "mistralai/Timestral-209M"}'

Notes:

• Zero-shot: no fev-bench data in the training data (trained_on_datasets = []).
• Past and known covariates are used. Categorical covariates are target-encoded per series, and static columns are not used.
• The forecaster defaults are used, and the context is the full history up to 32768 steps.
• Submission tests pass locally.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[shchur]

I would prefer to only show 1 model variant to avoid crowding the leaderboard and distorting the win rates. Can you please set hidden: true for the 5m, 28m and 123m models (so only 209m is shown)?

[adrienpetralia]

Done, thanks! Only 209M is shown now.

[shchur]

Thanks for the contribution and congrats on the release!