binary_increment: default compile width d=512

At d=256 the optimize=1 schedule needs 17 layers; from d=512 up it sits
at the 16-layer optimum (d_head and further widening don't move it).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: physicsrob

examples/binary_increment.py

@@ -35,7 +35,7 @@
 from torchwright.ops.swiglu.map_select import map_to_table, select
 from torchwright.ops.swiglu.sequence_ops import output_sequence, remove_leading_0s
 
-D_MODEL = 256
+D_MODEL = 512
 # Rotary width the graph is built against; must match the d_head it is
 # compiled at (the token-example harness compiles at this d_head).
 D_HEAD = 16