test: cover OTel replay and user callback contexts - #758

Add Python handlers and SAM mappings for OTel cases 21–24: completed-step replay, active context in user functions and callbacks, and invocation retry status. User spans use ordinary implicit parenting. The handlers exercise public APIs through initial execution, wait/resume, failure, overlapping parallel/map branches, polling, callback submitters, retry helpers and virtual contexts.

Shared requirements: aws/aws-durable-execution-conformance-tests#131, pinned consistently to 6182cee00ed33ce01e0579066fb31096f988817f. Case 24 accepts canonical RETRYING followed by SUCCEEDED, optionally with a valid PENDING invocation between them; unexpected statuses/phases remain failures. This PR includes focused runtime prerequisites #756 and #752; merge those and the shared requirements first. Python #737 is excluded.

The context correction isolates failed plugin setup and preserves token ownership, caller restoration and the original no-plugin path. Handler scopes require explicit concrete-class opt-in, leaving legacy helpers/properties/dynamic attributes untouched. Both OTel views preserve same-trace parents and baggage and bind their existing durable span when the worker parent is absent or unrelated. Provider API 1, dependency floors, registration lifetime and checkpoint/replay formats remain unchanged. Older supported cores retain their existing behavior.

Validation:

• All 83 core-package files match fix: propagate context into handler workers #756's tested tree: 1,778 core tests. The combined current-core OTel suite passes 404 tests; actual installed minimum core 2.0.0 passes 356 applicable tests. Repository type checks pass.
• OTel conformance package 65, normal conformance package 11, and workflow wiring 3 tests pass. Case 22 uses the real local runner and barriers to force sibling overlap, without repairing tracing context in user functions.
• The extended real SDK scope matrix covers span-binding plugins in both orders, same/unrelated/absent context, resumed success/failure, exact parents and cleanup. Four real installed core/plugin pairs pass 24 compatibility cases; current-core capability checks remain active and the legacy boundary is explicit.
• Concurrent fixture hook output is serialized into complete JSON lines, backed by a regression that fails before the fix. Only the read-only ADOT layer lookup gets bounded standard retries for observed throttling. Requirements, wait timings, test timeouts, assertions, IAM and workflow trust are unchanged.
• PR CI runs all 24 OTel cases in both views through ADOT/X-Ray and S3 against the exact SDK head and pinned requirements, with the failed-plus-uncovered gate enabled.

The earlier local tester duplicate wait-continuation behavior reproduces on unchanged main; deployed coverage and the original failure evidence are retained. Current-head cloud results and actual AI review generation/posting are verified separately.