perf.json cannot hold per-GOARCH allocation counts #6

perf.json holds one allocation count per case, but a count can depend on the architecture: an adaptive solver stores one state per step, and where the compiler fuses multiply-adds (arm64) a stiff problem takes a different number of steps than on amd64.

Witness: go-numerics/integrate, SolveIVP Radau on Van der Pol (mu = 1000): 353 allocations on linux/amd64, 350 on arm64; Radau with a Jacobian 349 and 345; BDF 596 and 595. A single perf.json fails CI mode on one of the two.

Interim in go-numerics/integrate: bench/perf_test.go selects docs/perf-arm64.json and docs/perf-table-arm64.md on arm64 through CIConfig.JSON and Table.

Request: let a report carry per-GOARCH allocation counts (or have RunCI pick perf-<goarch>.json when present), so no repository needs its own selection code.

Tracking: go-scikit/roadmap#27.