Supply Model Target retry policies from backend adapters

[adamtheturtle]

Shared Model Target helpers currently define the real service's retry criteria, attempt limit, and delays, and track whether the real backend is active. That makes a reusable contract suite own an adapter's failure policy.

Move that configuration into fixtures/real_model_target_retries.py. Adapters now supply a scoped tenacity.Retrying policy through retrying_model_target_requests. Shared helpers send requests once by default and retain the GET-only guard. Nested scopes restore the preceding policy, including after exceptions. The real adapter installs the same criteria, retry limit, delays, and final-outcome handling as before.

The existing retry tests use a copy of the real policy with wait_none(), removing the time.sleep monkeypatch. Two additional tests check installation through the public real Model Target setup fixture and restoration after a nested consumer policy raises. These checks use controlled send callbacks without account access or network calls.

Part of #3738 and #3737. This leaves the remaining shared-suite decoupling work open.

Validation:

• Full local suite excluding real-service and Docker-build cases: 2,016 passed, 839 skipped.
• Focused retry, timeout, and Model Target response-policy tests: 38 passed.
• All 2,853 existing collected cases and IDs preserved; two cases added.
• Independent consumer smoke: six passed with native adapter, real-policy, real-credential, settings, and Flask imports blocked. Its custom policy uses different retry criteria and a two-attempt limit, preserves both supported response types, and sends POST requests once.
• Shared retry helper, real retry policy module, and retry tests fully covered locally.
• Pre-commit, pre-push checks (including partition and all type checkers), changed-file Pylint, and diff checks passed.
• AST comparison confirms the moved real retry configuration and final-outcome handling are unchanged.

Docker and live-service validation remain with hosted CI.

Hosted CI is currently blocked by live Vuforia HTTP 429 responses carrying x-envoy-ratelimited: true. The changed retry-policy tests, mock tests, and lint checks passed. Rerunning failed jobs did not clear the throttle; an isolated retry of update_target/test_width.py, with no other active Test workflows in either VWS repository, also exhausted retries with the same response. Isolated failing job. The live-service checks need another run after the throttle clears.