Add etcd-backed durable cluster state

[halfcyan]

Summary

Store Odorobo's desired cluster state in etcd so manager restarts and cluster outages do not cause VM definitions or placement data to be lost.

Scope

• Add an etcd client abstraction behind a trait.
• Define key prefixes and serialized records for VM manifests, placement, node state, and operations.
• Add endpoint, TLS, and authentication configuration.
• Add connectivity/health diagnostics.

Tasks

• Select and add a compatible etcd client dependency.
• Define the storage trait used by the scheduler and reconciliation code.
• Define key prefixes for VM manifests, placement, node state, and operations.
• Define serialized record versions and migration behavior.
• Add endpoint, TLS, authentication, timeout, and retry configuration.
• Implement read/write/delete operations for VM state.
• Add connectivity and health diagnostics.
• Add mock-client or isolated-etcd tests.
• Document behavior when etcd is temporarily unavailable.

What this should do

• VM state can be written and read through the abstraction.
• Key layout and serialization are documented.
• Connection failures are surfaced clearly.
• Tests use a mock client or isolated etcd environment.
• No destructive behavior occurs when etcd is temporarily unavailable.

Rationale

Use etcd as the initial source of truth instead of making the DHT a prerequisite for the MVP.

Issue links

• Foundation for #101 and #107.
• Works with #98 and #99.