Add sampled divergence log to labkit rack shadow

What does this MR do and why?

Adds a temporary, sampled structured log line for divergent block decisions in the labkit rack shadow middleware. The existing gitlab_rate_limiter_labkit_rack_shadow_total counter labels each data point only by whichever side blocked, so it cannot distinguish an identity-taxonomy defect from window-phase skew. Each logged divergence now carries, side by side:

• both stacks' verdicts (labkit_blocked / rackattack_blocked)
• every matched labkit rule with its action, count, and limit (shows where labkit routed the request, including synthetic skip/bypass matches)
• every Rack::Attack throttle annotation with its discriminator, count, and limit (a throttle absent here means Rack::Attack did not count the request at all)
• the labkit identity facts (requester_type, requester_id, runner_id, remote_ip, path, method)

This is the first unchecked item in the diagnosis plan for the parity divergence classes blocking the per-cohort enforce gates: it separates the api over-blocking class from the git-LFS window-phase class mechanically, and pins the remaining identity-attribution asymmetry that metrics cannot distinguish.

Volume is operator-controlled: logging is gated by the new default-off ops flag log_labkit_rack_divergence, sampled per request via percentage-of-actors on Feature.current_request, and only divergences (never matches) are logged. With the flag off the only new work per divergence is one Feature.enabled? call.

References

• https://gitlab.com/gitlab-com/gl-infra/production-engineering/-/work_items/29362
• Rollout issue: #605893

How to set up and validate locally

1. Enable a shadow cohort flag and the log flag: Feature.enable(:rate_limiter_use_labkit_rack_cohort_2), Feature.enable(:log_labkit_rack_divergence) (or Feature.enable_percentage_of_actors(:log_labkit_rack_divergence, 1) for sampling).
2. Drive requests past one stack's limit and watch log/application_json.log for Labkit rack shadow divergence lines.