[ChrisW09]

Fixes #487

Problem

LSS fit() validated target support only for poisson/negativebinom (non-negative) and gamma/inversegamma/lognormal (strictly positive). For the other families:

• Fractional targets for the count families crashed mid-training with raw torch support errors, after preprocessing and model construction. So did negative zip targets and beta targets outside (0, 1). The previously fitted estimator was discarded on the way.
• Negative tweedie targets trained silently.

Fix

_validate_target_array (deeptab/models/base.py) now also checks the count families, tweedie and beta:

Family	Valid targets
poisson, negativebinom, zip	non-negative integers (integer-valued floats are fine)
tweedie	non-negative values
gamma, inversegamma, lognormal	strictly positive (unchanged)
beta	strictly inside (0, 1)

Violations raise DataError (family='…' requires … target values) during input validation, before any estimator state changes. This is the same path the positive-support families already used, so a failed call keeps the previous fitted model. y_val goes through the same check. Targets that were valid before are still accepted. The only behaviour change is that inputs which used to crash mid-training or train on invalid targets are now rejected up front.

Tests

• tests/core/test_exceptions.py:
  • count families reject negative and fractional targets and accept integer-valued ints and floats;
  • tweedie rejects negatives and accepts zeros;
  • beta rejects 0, 1 and values outside the interval, and accepts values inside it.
• tests/integration/test_estimator_input_validation.py: MLPLSS.fit raises DataError for each new case without building a data module or model.

19 of the 26 new cases fail on main; the 7 that pass are acceptance cases (valid targets that must keep working). All pass with this change.

dirichlet, whose targets must be positive proportions summing to one per row, is not checked here; it still fails during training on invalid targets.

Docs

docs/core_concepts/training_and_evaluation.md replaces the sentence that covered only the positive-support families with a table of the valid targets per family.

Checks

• just test: 3262 passed, 46 skipped, 28 xfailed; coverage 91.29%
• just docs: build succeeded (warnings treated as errors)
• just check: all hooks passed (ruff, prettier, pyright)