perf: compare OID arcs for equality before ordering them

[ilyam8]

Summary

Follow-up to the phase 6 re-evaluation. No behavior change.

oidCompare compares OIDs arc by arc. Since #56 it ordered each arc pair with cmp.Compare, which makes two
comparisons (<, then >), where the loop before made one inequality test; most arcs of the OIDs a walk compares are
equal, so every call paid for the extra comparison (+7.4% per call, measured against fcca7b0). Testing for equality
first and calling cmp.Compare only for an arc pair that differs keeps the cmp.Compare form and removes the cost.

A walk calls oidCompare once per value, so the effect on real walks is a few nanoseconds per value; it showed as
+9-11% in BenchmarkWalk, whose test agent finds each answer with a linear oidCompare scan.

Validation

• TestOidCompare and the walk goldens unchanged; six mutants of the changed lines (inverted test, swapped
  arguments, no test, constant results, a one-sided test) all fail the tests.
• BenchmarkWalk (8 interleaved rounds against docs: document the package, its known bugs, the README and the examples #57's merge): WalkAll -8.7%, BulkWalkAll -12.5%, back to their
  times before refactor: tidy the request engine and API #56; bytes and allocations unchanged.
• golangci-lint, race tests, FIPS 140-only, cross-builds, govulncheck: pass locally.