Fix MLX implementation of LocalDotProductSelfAttention that previously overwrote block_size

[lancelotblanchard]

The current MLX implementation of LocalDotProductSelfAttention does:

@property
@override
def block_size(self):
  return self._block_size_config

, which overwrites the block_size property of the class and breaks inference using step, which needs block_size=1. Instead, the JAX implementation does not overwrite the property. This PR fixes the MLX implementation to match the JAX one.

Additionally, this PR adds corresponding attention tests (self.assertEqual(layer.block_size, 1)) for the step function, and modifies DefaultSteppable to make sure properties are read from base class backend_sl.types.Steppable and not DefaultTestLayer.

[rryan]

Great catch -- this is an unfortunate name clash. block_size in LocalDotProductSelfAttention is unrelated to the block_size concept in the SequenceLayer API. I imagine this got mixed up in the initial MLX port.