Fix binary_cross_entropy docstring argument order (weights before with_logits)

[Pablosinyores]

The binary_cross_entropy Args block documents the parameters as inputs, targets, with_logits, weights, reduction, but the actual signature order is inputs, targets, weights, with_logits, reduction:

def binary_cross_entropy(
    inputs, targets,
    weights=None,          # 3rd positional
    with_logits=True,      # 4th positional
    reduction="mean",
)

A caller following the docstring order positionally, e.g. binary_cross_entropy(logits, targets, False) intending with_logits=False, instead binds weights=False. The weights is not None branch then does weights.shape on a bool and raises AttributeError: 'bool' object has no attribute 'shape'.

>>> nn.losses.binary_cross_entropy(mx.array([0.1,0.2]), mx.array([0,1]), False)
AttributeError: 'bool' object has no attribute 'shape'

This reorders the two Args entries so weights precedes with_logits, matching the signature. The sibling cross_entropy in the same file already documents weights in its correct signature position. Docstring-only change.