Add opt-in retry-with-backoff for rejected pushes - #1090

Why

When many independent ImageUpdateAutomation objects push to the same branch of one GitOps repo (e.g. one IUA per service/region, all on independent 5-minute reconcile timers), a multi-service release can cause many of them to try pushing around the same time. A concrete example observed in production: five regions all resolved a new image tag within 1-2 minutes of each other, but the resulting git commits landed 13-29 minutes apart, because several IUAs lost a non-fast-forward push race against others pushing to the same branch:

14:32:43  Warning  failed to update source: failed to push to remote: object not found
14:45:40  Normal   pushed commit 'a622711' — Release java-backend:2026.34.2

There is no in-process retry today — the next attempt only happens on the next scheduled reconcile, or via controller-runtime's per-item exponential-backoff requeue (750ms doubling, capped at 15min). Under heavy contention, a single IUA can lose several races in a row, each costing a full backoff cycle, which is how a sub-minute git operation turns into a 29-minute delay. This reproduces on every release with enough concurrent writers to one branch, and gets worse as more services onboard per-instance IUAs.

A retry-with-backoff loop that reuses the existing local clone (fetch + reset instead of a full re-clone) resolves a lost race in seconds instead of minutes, following the same shape as a standard retry_with_backoff(5, 2, ...) wrapper around a git pull --rebase && git push cycle.

What

• commitAndPushWithRetry (new internal/controller/push_retry.go) wraps CommitAndPush: on a push rejected specifically because another writer already advanced the branch (source.IsPushConflict), it waits (2s/4s/8s/16s backoff, 5 attempts total), fetches and hard-resets onto the new remote tip (SourceManager.RefreshToRemote, no full re-clone), re-applies policies against the refreshed tree, and retries the commit and push. Any other error, or exhaustion of all attempts, is returned unchanged — existing condition/error handling in reconcile() is untouched.
• The retry loop is bounded by min(.spec.interval / 2, 2 minutes) so contention on one branch can't block a reconcile worker indefinitely and starve other ImageUpdateAutomation objects sharing the pool.
• Depends on (*gogit.Client).FetchAndReset and git.ErrPushRejected from git/gogit: add FetchAndReset and ErrPushRejected for push-conflict recovery pkg#1289 (companion PR) — go-git has no rebase API, so this fetches + hard-resets instead of a real rebase, which reaches the same end state here since policies are re-applied fresh against the reset tree rather than blindly replaying a stale diff.
• New feature gate GitPushRetryOnConflict, disabled by default — opt-in, so this ships as an explicit choice rather than a behavior change for existing installs.

Testing

• internal/source: TestSourceManager_RefreshToRemote_RecoversFromPushConflict — forces a real non-fast-forward rejection via a competing out-of-band push, then proves recovery lands the new commit on top of the competing one (parent-hash check), not an overwrite.
• internal/controller: five tests covering recovery within N attempts, exhausting all attempts against a real conflict, a non-conflict error (connection failure) returning after exactly one attempt, retry-budget enforcement under a short .spec.interval, and the feature gate's off-by-default fallback path.
• Full go test ./... passes (envtest via make install-envtest), go vet/gofmt clean.

Notes for reviewers

• go.mod currently points github.com/fluxcd/pkg/git at git/gogit: add FetchAndReset and ErrPushRejected for push-conflict recovery pkg#1289's fork branch (via a resolved pseudo-version, not a local path) so this branch builds standalone. Will be switched to a tagged release version once git/gogit: add FetchAndReset and ErrPushRejected for push-conflict recovery pkg#1289 merges and releases, before this is marked ready for review.