Record timing per dispatch, not per operation

[cymbalrush]

The runtime measures a compile identifier -- a possibly-fused group of Core AI operations -- as a whole, so a duration belongs to the group rather than to any one member. It also reports each encoder twice; only one of the two measurements is kept now, since pooling them put two different quantities in one sample list.

• OperationTiming carries op_ids and one undivided measurement, and the summary lists every member rather than a representative, which collided.
• _timings is keyed on (odix_id, sorted op ids): delegate_id is a per-dispatch counter, and an unsorted key split one dispatch across rows.
• An untimed warmup iteration, since a first inference is not the steady state.
• operations_by_id exposes every operation in the function, so a caller can compute coverage on its own terms.
• Module timings are filed against the deepest module containing all of a dispatch's operations, and no module reports a total: fusion crosses module boundaries, so a total charges one module for a sibling's work.
• Removed total_duration_ns and the per-operation lookups, which presented a profiling span and a shared duration as latency and per-operation cost.

Adds timing_diff for comparing two runs. Op ids are positional, so an edit renumbers them and comparing by id compares unrelated work. op_id_alignment supplies the correspondence, and dispatches pair only when they cover the same operations; one that gained or lost operations is reported as resized, with the delta attributable to them, and anything else as present on one side only.