Add lifecycle tracing spans to instance delete path

[sjmiller609]

Summary

DELETE /instances/{id} shows a heavy latency tail in production (p50 ~300ms, p95 ~3.4s), but its trace spans are opaque: the delete path in lib/instances/delete.go runs the same teardown sequence as stop (graceful guest shutdown, hypervisor kill, network release, data removal) without any of the lifecycle step spans that stop.go already emits, so the slow phase is invisible in traces.

This wraps each blocking phase of deleteInstance in the existing startLifecycleSpan/startLifecycleStep helpers, mirroring stop.go:

• top-level instances.delete span with hypervisor and instance_state attributes
• cancel_compression_job
• graceful_guest_shutdown (marked as error on timeout, same as stop)
• force_kill_hypervisor
• release_network
• delete_instance_data

No behavior changes — instrumentation only.

Test plan

• go build ./lib/instances/ and go vet ./lib/instances/
• go test ./lib/instances/ -run TestDeleteInstance passes
• Full lib/instances suite not run to completion locally: Firecracker VM lifecycle tests require real hypervisor binaries/network and fail identically on unmodified main in this environment
• After deploy, verify delete step spans appear under /instances/{id} DELETE traces

🤖 Generated with Claude Code

---

Note

Low Risk
Observability-only changes around existing delete flow; no logic or API changes.

Overview
Adds OpenTelemetry lifecycle instrumentation to deleteInstance so DELETE latency can be broken down in traces, matching the pattern already used on stop/create.

A top-level instances.delete span records instance_id, hypervisor, and instance_state via enrichInstancesTrace, with the function signature updated to named retErr so finishInstancesSpan runs on exit. Child step spans wrap the blocking teardown phases: cancel_compression_job, graceful_guest_shutdown (failed attempts end the step with errGracefulShutdownFailed, same as stop), force_kill_hypervisor, release_network, and delete_instance_data.

Teardown logic, timeouts, and error handling are unchanged—only tracing hooks around existing calls.

^{Reviewed by Cursor Bugbot for commit 38dea8a. Bugbot is set up for automated code reviews on this repo. Configure here.}