fix: agent exits promptly after a short trace drains

[CodingInAVan]

Problem

gpufl trace --upload hung for ~60s at the very end of short traces (e.g. the tutorials) even though the upload UI already showed "complete." The launcher would eventually hard-kill the agent and print a misleading "late windows may need a post-hoc gpufl upload" warning.

Root cause

In exit-when-drained mode, GpuflAgent.awaitDrainThenExit only declared a session "drained" after it had observed the live .tmp/ window marker through its 1s poll (sawActive). A short trace creates and finalizes (removes) that marker between two polls, so sawActive never flipped — the drain loop spun forever. The launcher then waited out its full --agent-drain-ms cap (60s) and hard-killed the agent.

Fix

Gate the drain on a cumulative signal instead of the race-prone .tmp/ observation:

• TailerManager.hasStartedAnySession() — backed by the startedSessions keyset, which is added to on discovery and never cleared.
• Drain = started && activeTailers == 0 && !anyActiveSession, two consecutive clean polls.

The window-model LogTailer only decrements activeTailers after a window is fully sent, so there's no premature-exit risk. The agent now self-exits ~2–3s after the upload drains.

Refactor (folded in)

Centralized the agent's magic-number sleeps into a Delays utility with an interrupt-aware sleep() helper, replacing ad-hoc Thread.sleep/catch blocks across GpuflAgent, LogTailer, SessionWatcher, and TailerManager. signalSessionComplete also gains exponential backoff and slf4j logging.

Testing

• ./gradlew build green (unit tests + jacoco).
• Verified end-to-end: a short gpufl trace --upload now finishes immediately after "agent finished uploading" — no 60s hang.