fix: raise clear ValueError when add_batch receives None inputs

[YousefZahran1]

What

add_batch() crashed with a cryptic TypeError: object of type 'NoneType' has no len() when predictions or references was None. The error originated at line 518 (if len(column) > 0), was caught by the except (pa.ArrowInvalid, TypeError) handler, then immediately crashed again at line 523 (len(batch[c])) — producing a second, equally confusing traceback.

Why

This affects anyone following the HuggingFace Trainer tutorial whose compute_metrics function conditionally returns None (e.g. when evaluation data is partially unavailable). The root error message gives no indication which field was None or why.

Reproduce

import evaluate
acc = evaluate.load("accuracy")
acc.add_batch(predictions=None, references=[0, 1])
# Before fix → TypeError: object of type 'NoneType' has no len()
# After fix  → ValueError: Batch inputs contain None for the following fields: ['predictions']. ...

Fix

Validate the batch dict immediately after construction and raise a descriptive ValueError that names the offending fields and tells the user to check their compute_metrics return value. The change is 6 lines.

Testing

import evaluate
acc = evaluate.load("accuracy")
try:
    acc.add_batch(predictions=None, references=[0, 1])
except ValueError as e:
    print(e)
# Batch inputs contain None for the following fields: ['predictions'].
# All inputs must be non-None sequences (lists, arrays, or tensors).
# Check that your compute_metrics function passes non-None predictions and references.

Fixes #668