Fix: use math.log for Python float bounds in hyperparameter_sampling

[jasherb]

Problem

test_distribution_normalizers has been failing since f04d3ca ("sync"):

TypeError: log(): argument 'input' (position 1) must be Tensor, not float

In pfns/priors/hyperparameter_sampling.py, self.lower and self.upper
are Python float attributes of the dataclass, but f04d3ca changed
math.log(...) to torch.log(...) for both, in three methods.
torch.log doesn't accept Python scalars, so it raises immediately on
any call into normalize() (or the new unnormalize()) with log=True.

Fix

Revert the scalar-side calls to math.log in:

• UniformFloatDistConfig.normalize
• UniformFloatDistConfig.unnormalize
• PowerUniformFloatDistConfig.unnormalize

The tensor-side torch.log(value) calls stay as they were.

Verification

Before:
1 failed (test_distribution_normalizers), 30 passed

After:
pytest tests/ -ra --ignore=tests/priors/test_convert_prior_to_x_only_format.py
31 passed

(The ignored file has separate pre-existing failures also introduced in
f04d3ca — happy to file a follow-up PR if useful.)