fix: bound the eventlet Glance API worker count by default

[berendt]

Problem

Running a sustained image-churn load against a ControlPlane-managed
Glance (release 2025.2) crash-loops the API pod. The container is
OOMKilled (exit 137) within seconds of taking load and never becomes
ready, so clients see connection resets and 5xx while it flaps.

Root cause

Below release 2026.1 Glance launches the eventlet glance-api
server, whose worker count comes from [DEFAULT] workers in the
config. The operator rendered that key only when
spec.apiServer.workers was set explicitly; when it was nil the
eventlet server fell back to its own default of one worker per host
CPU. That fallback ignores the pod's CPU limit, so on a multi-core
node Glance forks one worker per core (e.g. 6 workers + master) into a
container capped at 512Mi and is OOMKilled under concurrent uploads.

This is a parity gap: the uWSGI launch mode (2026.1+) and keystone
already pin a bounded process count (DefaultUWSGIProcesses = 2); only
the eventlet path lacked the equivalent — even though the
spec.apiServer.workers godoc already promised the operator applies a
hardcoded default when the field is nil.

Fix

• Add DefaultEventletWorkers = 2 next to DefaultUWSGIProcesses as
  the single source of truth.
• Render [DEFAULT] workers in eventlet mode whenever the field is
  unset. An explicit value still wins, and the uWSGI mode is unchanged
  (the key stays inert there and is emitted only when set).

The worker count — and thus the memory footprint — is now
deterministic regardless of node size.

Testing

• New table test TestOperatorDefaults_EventletWorkers covers the
  launch-mode-conditional rendering (eventlet unset → default,
  eventlet explicit wins, uWSGI unset → absent, uWSGI explicit → inert
  render).
• Full operators/glance module tests, go vet, and gofmt pass.
• CRD description for spec.apiServer.workers regenerated and the Helm
  chart copy synced (make verify-crd-sync passes).

Verified end-to-end on a local kind ControlPlane: pinning the worker
count drops the pod from 7 processes to 3 and steady-state memory from
the 512Mi ceiling to ~130Mi, ending the OOM crash-loop.

🤖 Generated with Claude Code