Report debugging findings by the module and line they came from

[cymbalrush]

graph_diff answers which node became which, plan_diff which op id moved device, intermediates which op id diverged. None of those is the question a reader asked. A node id or an op id names nothing they wrote, and a report over a real model is not read at all unless it is grouped -- so every tool here stopped one step short of the answer, and each stopped in its own way.

This adds the step, once, and points the three reports at it.

modules.py -- the rollup, shared rather than per-tool.

build_module_tree groups any per-operation finding into the module instance tree it came from, generalising the insert-or-create walk that was inline in BenchmarkResult.get_module_timings. attributions_by_op_id composes the two existing lookups, which deliberately disagree on how to filter source locations: provenance must not trim frames or a model of stock nn modules collapses to one path, while a source line must trim them or it names a line inside torch.

Instances, not types. Block$1 and Block$3 are different answers, and an operation name cannot tell them apart: a model with twelve identical blocks has twelve identical linear operations, and "the slow one is a linear" is not a finding.

changes.py -- compute_changes, built on that rollup.

Three things stand between a correspondence and "7 operations changed, 6 of them in Block$3, at model.py:41". Each is a measured defect in the naive version rather than a refinement:

• A difference lands on a value node as readily as on an operation. Reported as-is, a rewiring inside a callee body says nothing was modified. Resolved through responsible_op.
• Most differing pairs are consequences of one edit, not edits: operands are other operations, so inserting one changes every neighbour's slots. _is_independent_change classifies each disagreeing slot by what happened to its endpoints and keeps only the two transitions nothing else reports -- a rewiring and a substitution. Chawathe et al. (1996) minimality.
• Which instance of an interchangeable operation is left unmapped is arbitrary. Unaided, adding a third block reports its extra constant against Block$1, and a reader concludes something changed in a block that did not. _relocate_to_new_modules breaks the tie on module scope, which the matcher never looked at. Additions and removals only: a modification names a pair and is interchangeable with nothing, so re-seating one end is simply false.

plan_diff and intermediates -- the same rollup, so "which layer" is one call whatever was measured.