fix: correct Tweedie deviance formula for power not in {0, 1}

[AlejandroCoronadoN]

Problem

For a general Tweedie power (and for the power == 2 gamma case), the factor
of 2 was only applied to the first deviance term:

2 * ( y^(2-p) / ((1-p)(2-p)) )   # only this term is doubled
- ( y * m^(1-p) / (1-p) )        # not doubled
+ ( m^(2-p) / (2-p) )            # not doubled

The unit Tweedie deviance doubles all three terms. Because of this, the result
was wrong for every power other than 0 and 1, and even negative for
1 < power < 2 (a deviance can never be negative).

power=1.5: tweedie_deviance = -6.67   sklearn mean_tweedie_deviance = 0.066
power=2.0: tweedie_deviance =  0.069   sklearn                       = 0.045

The bug went unnoticed because the loss comparison test only covered powers 0,
1 and 2, and its own reference for power == 2 repeated the same mistake.

Fix

Distribute the 2 * over all three terms in both the general-power branch and
the power == 2 branch.

Verification

After the fix, tweedie_deviance matches sklearn.metrics.mean_tweedie_deviance
for powers 0, 1, 1.2, 1.5, 1.8, 2.0, 2.5 and 3.0 (max abs diff < 1e-6). The test
reference for power == 2 is corrected to the same formula, and general-power
cases (1.5 and 2.5) are added to test_loss, which fail without the code change
and pass with it. The full test_losses.py suite (106 tests, pandas and polars)
passes.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ MMenchero
❌ AlejandroCoronadoN
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[MMenchero]

Hi @AlejandroCoronadoN thanks for this fix! I just updated the documentation to show the correct formula.

Can you sign the license so we can merge your PR? We would like to make a release asap, including your fix.