MixtureSampler accepts all-zero weights inconsistently

[mmshad]

MixtureSampler disagrees with itself on all-zero input. The two
normalization branches (selected by temperature) behave differently:

• temperature == 1.0: divides by sum(weights) and raises
  ZeroDivisionError mid-run.
• temperature != 1.0: clamps each weight with max(w, 1e-12) and silently
  produces uniform sampling.

Both are reachable from the phase-annealing path in scripts/train.py.
TrainingPhase.dataset_weights validates >= 0, not > 0, to intentionally
allow dropping individual datasets. A user who zeroes out every dataset in
a phase transition gets a mid-run crash on one branch and silent
training-quality drift on the other.

Fix

Add _validate_weights(weights, context) in kempnerforge/data/sampler.py
and call it from both MixtureSampler.__init__ and
MixtureSampler.update_weights. The helper rejects:

• empty weight lists,
• any negative entry,
• all-zero lists (sum <= 0),

with a ValueError that prints the offending list and the call site.

The TrainingPhase schema-level validator stays permissive because it
cannot know the effective mixture at config-parse time (phases only merge
at runtime). The sampler-level gate is where the hard check belongs.

Coverage

tests/unit/test_mixing.py::TestMixtureSamplerZeroWeights adds 8 tests
covering both constructors, both temperature branches, negative weights,
empty list, and a positive case where a single dataset is zeroed out but
at least one stays positive.