fix: check the Tweedie deviance domain for each power

[Raashish Aggarwal (raashish1601)]

tweedie_deviance checks that every prediction is strictly positive, whatever the power. With power=0 the deviance is just the squared error, so zero or negative predictions are valid there, but they raise:

df = pd.DataFrame({"unique_id": ["a"] * 3, "y": [1.0, -2.0, 0.5], "model": [0.5, -1.0, 0.0]})
tweedie_deviance(df, ["model"], power=0)
# ValueError: All predictions must be strictly positive for Tweedie deviance.

The opposite happens for 1 <= power < 2: a negative target is not checked, so it is silently scored. With power=1.5 the general branch clips the target at 0 and returns a negative deviance (-0.70 for y=[1, -2, 0.5], model=[0.5, 1, 2]), and with power=1 the log term is dropped for that row.

The checks now follow the domain of each power, the same as sklearn.metrics.mean_tweedie_deviance:

• power=0: any target and prediction.
• 1 <= power < 2: target must be non-negative (new check), prediction strictly positive.
• power >= 2: target and prediction strictly positive (unchanged).

Tests: added one test that power=0 accepts non-positive predictions and matches the mean squared error (pandas and polars), and one that power 1 and 1.5 reject a negative target. Both fail on main. tests/test_losses.py passes locally, and ruff format and mypy pass on the changed files.

[CLAassistant]

All committers have signed the CLA.