Count of correct top-1 predictions as an int64 0-d Tensor.

Equivalent to (logits.argmax(dim=dim) == target).long().sum() — written as a single function so the four-op chain in user code collapses to one Python wrap. The complementary primitive to accuracy: where accuracy divides by total count eagerly, correct_count keeps the count integer so the user can accumulate across batches and divide once at the end of the epoch.