fix(daemon): survive relay disconnect edge cases

[glittercowboy]

TL;DR

Daemon reconnect Hello frames now include active task IDs, handler-level relay errors no longer tear down the entire connection, interactive waits time out cleanly, duplicate task IDs are ignored, and temp uploads/image goroutines clean up correctly.

Why it matters

This keeps in-flight tasks alive across reconnects, prevents one bad browser or filesystem request from disconnecting all sessions, avoids duplicate execution, and closes cleanup leaks around temp files and upload goroutines.

How it works now

• Hello includes ActiveTasks
• readPump logs handler errors and continues
• actor sequence numbers use atomic increments
• permission and question waits time out and surface TaskError
• task-scoped image uploads use the task context
• duplicate task IDs are ignored per actor
• downloaded image temp files are removed after executor exit

Verification

• go test ./internal/relay ./internal/session ./internal/loop ./internal/claude
• go test -race ./...

Landing / post-merge

• Merge after the relay PR lands and the relay deploy is healthy.
• Daemon changes are backward-compatible, but the relay queue-state fix should ship first.