fix(ci): ride out transient network flakes instead of failing the run

[mhenrixon]

Summary

The three red CI runs on dash (merges of #22, #23, #24) were all infrastructure, none were test regressions:

Run	Symptom	Root cause
#24 merge	bundle exit 17 in Install Ruby (rails_edge, Ruby 4.0)	rubygems.org SSL reset mid-resolution — every job resolves from scratch since Gemfile.lock is removed
#23 merge	every MainTest failed at docker compose build	Docker Hub dial tcp i/o timeout pulling base images; build_images_once had no retry, so each test re-hit the outage
#22 merge	steps ended with conclusion: null	runner died mid-job (nothing to fix repo-side)

Changes:

• test/integration/integration_test.rb — build_images_once retries docker compose build up to 3 attempts with 15s/30s backoff, matching the compose_up_with_retry idiom right above it
• .github/workflows/ci.yml — BUNDLE_RETRY: "6" on the tests job; guard the always() disk-usage step so du /mnt/docker no longer paints a second misleading failure when an earlier step is what actually failed

Test plan

• CI green on this PR (actionlint + zizmor validate the workflow edits)
• Next transient Hub blip during integration build shows a "retrying" line instead of failing the suite
• An early job failure no longer flags Check disk usage as a second failed step

Deviations & judgment calls

• Branched off dash, not main — deliberate exception to the feature-branch rule: all touched files are fork-owned and heavily diverged from upstream (ci.yml matrix, integration harness), so a main-rooted branch would be pure conflict noise and this change is not upstream-PR-able anyway.
• Retries won't save a sustained outage — the feat(app): expose kamal-proxy rollout as kamal app rollout deploy/set/stop #23-merge Hub outage lasted 4+ minutes; 3 attempts with backoff rides out blips, not incidents. Re-run stays the answer for those, but a blip no longer redlines the suite.
• No registry mirror for host builds — the compose stack's hub-cache can't serve the host daemon's base-image pulls without a chicken-and-egg bootstrap; not worth the complexity for the observed failure rate.
• Nothing fixed for the conclusion: null run — that's a GitHub runner death; no repo-side change can prevent it.