[BUG] Classifier labels are never encoded to 0..K-1 — crashes on string labels, silently wrong on non-contiguous integers #409

Describe the bug
Class labels are never encoded to 0..K-1 anywhere in the classifier fit path. SklearnBaseClassifier.fit stores self.classes_ = np.unique(y) (deeptab/models/classifier_base.py:294) and predict() maps argmax indices back through classes_ (classifier_base.py:386-388), but the raw y values are passed straight through _FitMixin._build_model → TabularDataModule.preprocess_data → torch.tensor(self.y_train, dtype=torch.long) (deeptab/data/datamodule.py:275-287) and used directly as CrossEntropy class indices / BCE targets. No LabelEncoder, np.unique(..., return_inverse=True), or searchsorted exists in the package.

Three concrete failure modes:

1. String labels crash — the raw y also reaches preprocessor.fit(X_train, y_train), and pretab's PLE decision-tree binning calls float(y):
   ValueError: could not convert string to float: 'yes'. sklearn classifiers must accept string labels.
2. Non-contiguous multiclass int labels (e.g. {10, 20, 30}): CE receives target 10 with 3 logits → IndexError: Target 10 is out of bounds on CPU. On MPS the fit completes silently (no bounds check) and the model trains on garbage.
3. Binary labels other than {0,1} (e.g. {5, 7}): BCEWithLogitsLoss silently trains against targets 5.0/7.0. On perfectly separable data the model reaches 47.75% accuracy vs 100% with {0,1} labels. No warning, no error.

To Reproduce

import numpy as np, pandas as pd
from deeptab.models import MLPClassifier

rng = np.random.RandomState(0)
X = pd.DataFrame({"num1": rng.randn(400)})

# (1) string labels -> ValueError in preprocessor
y = np.where(X["num1"] > 0, "yes", "no")
MLPClassifier().fit(X, y, max_epochs=1)

# (2) non-contiguous multiclass -> IndexError on CPU, silent garbage on MPS
y = rng.choice([10, 20, 30], 400)
MLPClassifier().fit(X, y, max_epochs=1, accelerator="cpu")

# (3) binary {5,7} -> silently broken training
y = np.where(X["num1"] > 0, 7, 5)
clf = MLPClassifier()
clf.fit(X, y, max_epochs=30, batch_size=32, accelerator="cpu")
print((clf.predict(X) == y).mean())   # ~0.48 on perfectly separable data; 1.0 with labels {0,1}

Expected behavior
fit should encode labels to contiguous indices (and pass the encoded y to the preprocessor and datamodule); predict/predict_proba already assume this mapping exists. Arbitrary string/integer labels should round-trip like any sklearn classifier.

Screenshots
n/a

Desktop (please complete the following information):

• OS: macOS (Darwin 25.5.0, arm64)
• Python version: 3.11.15
• deeptab Version: 2.0.0 (main @ 4e6a359)

Additional context
torch 2.9.1, lightning 2.6.5, scikit-learn 1.9.0, numpy 2.4.6. The existing test suite always uses 0..K-1 integer labels, which is why this never surfaces. Case (3) is the most dangerous: silent, and the MPS variant of case (2) is silent too. Found during a systematic bug-hunt of v2.0.0.