Share benign-race classifier; match landed code-only form

[LandynDev]

Log-hygiene sweep of the benign-race handling around permissionless cranks. No behavior change to flow control; one log-level fix.

• One classifier. benign_marker in solana/client.py replaces five copies of the any(m in str(e)) predicate (swap loop x3, weights vote, CLI crank). Returns the matched error NAME.
• Close the code-only gap. A benign race that lands on-chain stringifies as {'Custom': N} with no name. CLI already matched both forms; the validator matched names only, so a landed benign race logged a false-alarm ERROR. All sites now match both via a single name->code map.
• Terse benign logs. Benign no-ops log the marker name (e.g. no-op (PoolNotClosed)) instead of the full ~1KB RPC simulation blob.
• Accurate label. resolve_pool no-op (peer won the race) dropped - PoolNotClosed/SeedSlotNotYetProduced mean the validator was early, not that a peer won.

Tests: classifier test covers name form, landed code-only form, and non-benign passthrough. Full suite passes (795).