Foreach grouping: isomorphic elementwise chains as torch._foreach_* calls

New emission-level planning pass _plan_foreach (FOREACH_GROUPING flag,
plan/suppress idiom like SDPA/addmm): families of shape-isomorphic
MapNode/LinearNode/pure-Hadamard-EinsumNode chains — the 53 per-parameter
AdamW update chains — key by a shape-erased structural signature
(op + scalar params / ordered weights / broadcast-multiply steps, operand
roles recursive) and emit at the family anchor as one OUT-OF-PLACE
torch._foreach_* call per chain step, tuple-unpacked onto the members'
pre-assigned t-names so output collection, the regex DLE and EMIT_DEL
liveness keep working textually. A verify-and-demote fixpoint guarantees
deferral never moves a producer past a consumer. Scalars enter as python
floats; identity/0-dim alignment patterns are rank-erased so 2-D weight
chains and 1-D bias chains share a family. Emission mirrors the singleton
ATen ops exactly (mul.Scalar then add, never add(alpha=)), so the flag
A/B is bit-identical.

Status: correct and partial — 27 grouped calls on the nano step (sqrt
54->30, reciprocal 62->38); the remaining singletons demote through
cross-family anchor inversions and sub-minimum families (census in the
session log). Window-based family scheduling to recover them is filed.
Implementation largely by a background agent against the workflow-judge
design; finished and gated here.

Suite 884 passed; demo trains to 1.000; pyright clean on this file.

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

Author: thomasahle