Fix bring-up bugs (transformers >=5 / datasets >=4 / py3.14)

[matteius]

While bringing up the DNABERT-Epi training pipeline against current
PyPI deps (torch 2.11+, transformers 5.x, datasets 4.x) on Python 3.14
I hit a small set of upstream bugs that block end-to-end execution.
Each is independently fixable, none changes algorithmic behaviour.

Bugs fixed

1. pair_finetuning_dnabert.py

Bug	Symptom	Fix
dataset_dict[\"sgRNA\"] (capital R)	KeyError on first dataset — data_loader.load_dataset_information returns the key as lowercase `"sgrna"` (data_loader.py:168)	use \"sgrna\"
sgrna_seqs[dataset_name].extend(...)	TypeError: list indices must be integers — sgrna_seqs is [] not {}	direct .extend() on the lists
range((n_samples / n_sgrna) // 6)	TypeError: 'float' object cannot be interpreted as an integer — true-division returns float on Py3	wrap in int(...)
Trainer(..., tokenizer=self.tokenizer)	TypeError: unexpected kwarg — renamed in transformers 5	processing_class=self.tokenizer

2. data_loader.py — BalancedSampler.__init__

isinstance(dataset, Dataset) checks torch.utils.data.Dataset, but
the actual object passed in is datasets.arrow_dataset.Dataset
(unrelated classes). It then falls to the else-branch which expects
torch.Tensor from dataset[\"labels\"], but datasets >= 4 returns a
Column instead — .tolist() never runs and self.labels ends up
as a sequence of 0-d Tensors. The defaultdict buckets each tensor by
id() so the 0 in label_to_indices check fails even when both
classes are present (`ValueError: Dataset must contain both classes 0 and 1`).

Fix: more permissive normalisation cascade so self.labels is
always a list of plain Python ints.

3. dnabert_module.py — test_scratch + test_transfer

test_dataset[\"labels\"].numpy() → AttributeError: 'Column' object has no attribute 'numpy'. Same datasets 4.x root cause. Replaced
with np.asarray(list(...)).

4. dnabert2_module.py (new file)

run_preprocess.py imports models.dnabert2_module unconditionally,
but the file is missing from the repo (ImportError before any work
runs). The module is never actually called on the DNABERT (no-2)
code path; this stub keeps the import resolvable.

End-to-end verification

After these patches, on Lazzarotto 2020 GUIDE-seq fold 0 (scratch
mode, RTX 4070, batch 128, 8 epochs):

• pair-finetune converges to train_loss 0.060
• off-target classifier converges to train_loss 0.0028
• checkpoint saved successfully (fold0_iter0.pth, 82 MB)
• downstream evaluation runs (was the path that hit bug 3)

Known not-fixed-here

Bringing up on a fresh host also requires config.yaml to declare
two empty top-level dicts:

```yaml
dataset_name: {}
model_info: {}
```

Every entry-point writes into them without first creating the
parent (e.g. run_preprocess.py:46-47, run_model.py:54-56,
pair_finetuning_dnabert.py:250). I left this out of the PR because
it's a config-schema change rather than a code fix and may be worth
treating differently (e.g. defensive setdefault in
check_set.CheckConfig).