[CoR] Add durable polling, output capture, and cleanup for long deployments

[nturinski]

Problem

Long-running deployment commands can exceed execution windows and emit enough output to obscure status. Although some deployment paths suppress output or poll Azure operations, behavior is not consistent across channels. A timed-out command can make final status and cleanup harder to guarantee.

Proposed work

• Define channel-specific start-and-poll flows instead of relying on one long streaming command:
  • ARM/Bicep: start asynchronously and poll deployment state; capture failed operations.
  • App Service: use asynchronous deployment and poll deployment status.
  • ACR: suppress streaming logs and poll build state.
  • Flex Functions and Static Web Apps: document and implement their supported status mechanisms.
• Define command-specific time budgets and terminal-state handling.
• Persist sanitized command output under .copilot-azure/sessions/{id}/command-results/.
• Extend deployment audit entries with start/end timestamps, duration, terminal status, and output-file path.
• Keep chat output concise: progress state plus a short failure tail and a link/path to detailed output.
• Run cleanup and deploy-result.json finalization separately from timeout-prone deployment commands.
• Ensure temporary network access and credentials are reconciled idempotently after failure or timeout.
• Avoid command wrappers that obscure the actual process exit code.

Likely implementation surfaces include:

• resources/agents/azure-deploy/deploy/references/deploy-safety.md
• resources/agents/azure-deploy/deploy/references/deploy-checklist-template.md
• Channel-specific deployment references and deployment result/audit schemas

Acceptance criteria

• Each supported deployment channel has a documented start, poll, timeout, and terminal-state flow.
• Full sanitized diagnostics are written to a session-scoped file rather than streamed into chat.
• Timeout or cancellation still produces a terminal deploy-result.json state.
• Cleanup runs independently and removes/reconciles temporary access even after command failure.
• Audit data identifies duration, outcome, and diagnostic output location for every long operation.
• Tests or fixtures cover success, failure, timeout, polling interruption, and cleanup retry.