security: fixes for #90, #91, #93; agent-side mitigation for #92

[paulgnz]

agentescrow (no ABI change)

• [security] agentescrow: createjob + cancel loop grows contract RAM without bound (openjobs cap and cleanstale bypassed) #90: cleanstale also removes refunded jobs (state 7, escrow settled) older than
  stale_after, so a createjob+cancel loop no longer leaves permanent rows
• agentescrow: dispute→arbitrate settlement skips the platform fee entirely (approve = 1%, arbitrate = 0%) #93: arbitrate applies the platform fee to the agent's share, as approve does;
  still waived when the owner resolves as fallback

agentvalid (additive: tables openchals, chalseq)

• [security] agentvalid: challenge() rows are free, uncapped per validation, and survive every non-owner cleanup path #91: at most 3 unfunded challenges per challenger; cancelchal and expireunfund
  delete the row instead of flagging it; challenge ids come from a monotonic
  counter so a deleted id is never reused

openclaw

• agentescrow: client + third-party arbitrator collusion bypasses the conflict-of-interest guard (full escrow rug via arbitrate) #92: xpr_accept_job and xpr_submit_bid refuse jobs whose arbitrator is not in
  TRUSTED_ARBITRATORS (default protonnz); jobs with no arbitrator are allowed

Closes #90, #91, #93 once deployed (will close manually after mainnet hash verification).

Tests: agentescrow 280 passing (5 new, from the #90 PoC plus fee cases), agentvalid 53 passing (4 new, from the #91 PoC), openclaw arbitrator-guard tests (4).
ABI diff against mainnet: agentescrow unchanged; agentvalid adds tables openchals and chalseq only.