Refactor GPU endpoint lifecycle operations

[wang-tong0]

Summary

• replace the unrestricted af db set-endpoint command with af db register-static-endpoint
• stage static endpoints as inactive, preserve assignment/provider runtime fields, reject autoscaler-owned names, probe health, then activate with a conditional write
• serialize manual GPU add, replace, and remove requests through the running autoscaler daemon so normal scale ticks cannot race operator actions
• split af gpu add-endpoint from replacement and remove the ambiguous --keep-old path
• harden tunnel ownership checks, removed-endpoint cleanup, and autoscaler config metadata preservation

Why

The generic database endpoint command could overwrite scheduler assignment and autoscaler ownership fields. Manual GPU lifecycle commands also performed provider work outside the daemon, allowing them to race normal autoscaling and create duplicate or stale endpoint state. This change gives static endpoints and autoscaled endpoints separate ownership paths and makes the daemon the single executor for autoscaled lifecycle operations.

Safety

• static registrations cannot modify autoscale_managed=True rows
• runtime assignment and provider metadata are preserved during static config updates
• activation happens only after the configured SSH/Docker/GPU or Targon API preflight succeeds
• conditional writes reject ownership changes between staging, probing, and activation
• manual autoscaler requests use a single leased request, heartbeat, durable result, and cleanup path

Validation

• combined endpoint lifecycle and scheduler regression suite (347 passed)
• pytest -q (824 passed)
• ruff check on changed Python files; existing E402 command-registration imports in affine/cli/main.py ignored
• git diff --check origin/main...HEAD