fix(worker-ops): preserve queue tracking and prevent scheduler recovery stalls

[liobrasil]

Summary

This PR fixes three worker-ops reliability issues in FlowYieldVaultsEVMWorkerOps.

Commit Breakdown

1. 5350c4c — recover failed workers without scheduler stalls

• Prevents scheduler stalls by running failed-worker recovery before capacity gating only when capacity is saturated.
• Recomputes in-flight count after recovery so stale tracked entries do not block progress.

2. a5db0b3 — retain tracked requests when fail-marking fails

• stopAll(): removes scheduledRequests entries only if markRequestAsFailed(...) succeeds.
• _checkForFailedWorkerRequests(): keeps entries tracked when fail-marking fails, so recovery can retry instead of losing queue ownership.

3. 412dac2 — keep tracking on request lookup failures

• WorkerHandler.executeTransaction() now:
  • removes tracking on normal successful processing;
  • if lookup fails but tracked payload exists, attempts markRequestAsFailed(...) with tracked request data;
  • removes tracking only on successful fail-marking; otherwise retains tracking for scheduler recovery.

Why

Before these fixes, requests could remain PROCESSING on EVM while Cadence-side tracking was dropped, and scheduler recovery/capacity ordering could stall invalid/failing request handling.

Scope

• Cadence only: cadence/contracts/FlowYieldVaultsEVMWorkerOps.cdc