[Far0n]

Hi again,

Following up on #201, we'd like to submit TimeH2O-v1 to fev-bench.
It's a 108.8M-parameter time-series foundation model that runs through the h2ofm package.

On the full benchmark it gets a SQL skill score of 50.8 and a MASE skill score of 40.0, with a median runtime of 2.8 s per 100 series on a single RTX 5090.

This PR adds:

• models/timeh2o-v1/: the wrapper and requirements.txt (h2ofm==0.43.0)
• benchmarks/fev_bench/results/timeh2o-v1.csv: results for all 100 tasks
• an entry in models.yaml

To reproduce:

uv run python models/evaluate.py -m timeh2o-v1 -b benchmarks/fev_bench/tasks.yaml

A few notes:

• This PR currently sets hidden: true on the timeh2o-v1-tiny entry.
• The weights are gated on Hugging Face, so access has to be requested on the model page before running.
• The model's training data has no overlap with the fev-bench datasets.
• h2ofm requires Python 3.12 or newer (we used 3.14).

The submission tests pass locally.

Thanks again!

Cheers,
Far0n