Add background mode for shell commands so long runs don't hold steers

[TheGreatAxios]

Summary

• run_shell accepts background: true: shell-guard starts the command in a detached process group via a bounded registry (8 running, 8 completed ring) and returns {shell_id, status: "running"} immediately, so the turn ends and queued steers land while builds, test suites, and dev servers run
• On process exit the result (exit status, timed-out marker, ~2KB output preview, tool-output:/// spill URI when truncated) is delivered to the orchestrator on a later turn via the existing continuation channel, wired in all three loop hosts (TUI, exec, sub-agent); the message is system-originated and never counts as operator input
• New shell_collect tool ({shell_id, action: "collect"|"cancel", wait_ms?}) retrieves status/output or kills the process group; timeout kills the group with exit 124 + timed_out; background runs never mutate the retained shell cwd; the permission chain still evaluates the start call, so a denied command spawns nothing; unwired sessions fail closed; foreground behavior is unchanged when background is unset
• Tool watchdog exempts background starts and shell_collect; toolset dispose kills every live background process group before the posix teardown (/clear, interrupt, reload)
• Prompts, docs/ARCHITECTURE.md, and docs/IMPLEMENTATION.md updated

Verification

• bun run check passes (lint, typecheck, build, full test suite under the projects-dir guard: 6302 pass, 0 fail)
• New tests pin: immediate handle return, later-turn exit delivery (system-originated message shape), cancel and timeout process-group kills, ring eviction, registry/toolset dispose teardown, permission deny spawns nothing, foreground unchanged, background cd not retained

Fixes CL-7517

[linear-code]

CL-7517