Exclude managed Docker pauses from execution budgets

[Qian-Cheng-nju]

Docker freezes currently consume runner and checker wall-clock budgets during quota recovery. Add tlaps-bench pause-container and resume-container, backed by a durable host-owned ledger mounted read-only in the container. Use active elapsed time for runner deadlines and official checker subprocess waits, preserving separate wall/pause metrics, interrupted-check budgets, and retry accounting.

Claude Code benchmark children also receive CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS=0, so a freeze cannot spend the CLI's separate ten-minute background-wait ceiling and terminate a session before tool feedback arrives. The existing runner timeout still applies; a configured zero timeout remains unlimited. This setting is isolated to the child environment.

This applies to new Linux-hosted runs with an updated checker image and a quota guard using the new commands. Existing runs remain unchanged. Arbitrary shell timeout commands and internal tool timers retain their existing semantics; checks that depend on those timers must be paused between executions. Backends with an unadapted absolute deadline reject managed pauses. Non-Linux hosts retain their existing Docker behavior. See docs/PAUSING.md for usage and recovery boundaries.