weight_writer: fix d_head error message citing removed compile_game

The Attn d_head shape assertion's error message told the user to 'pass
d_head>=N or let compile_game auto-size it.' compile_game lived in the
deleted torchwright/doom/compile.py and isn't a callable anyone in this
repo can reach.

Replaced with 'Pass d_head>=N to compile_headless()' — compile_headless
is the actual public entrypoint and accepts the d_head kwarg directly.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: physicsrob

torchwright/compiler/forward/weight_writer.py

@@ -200,7 +200,7 @@ def _write_compute_attn(attn, op: AttnHeadOp, rmap: ResidualStreamMap):
     assert layer_d_head >= node.d_qk, (
         f"d_head={layer_d_head} is too small for Attn node "
         f"'{node.name}' which needs d_qk={node.d_qk} for Q/K. "
-        f"Use d_head>={node.d_qk} or let compile_game auto-size it."
+        f"Pass d_head>={node.d_qk} to compile_headless()."
     )
 
     # Q/K are shared across all V/O chunk heads, padded to layer d_head