fix(evaluation): expose shared span helpers and evaluator level lookup

[jariy17]

Consumers reading evaluation spans currently need private SDK helpers or copies of their implementation. This exposes is_tool_span, tool_span_ids, and trace_ids through bedrock_agentcore.evaluation.spans and the evaluation package, and routes both EvaluationClient and OnDemandEvaluationDatasetRunner through the shared implementation.

EvaluationClient.get_evaluator_level() also becomes public, reusing its existing control plane client and cache. Its existing SESSION fallback and caching behavior are preserved. The on-demand runner keeps its existing error propagation and locking behavior.

Related to #663 (span helpers and evaluator-level lookup). Includes usage examples and tests for all three span conventions, malformed attributes, missing IDs, ordering, duplicates, trace filtering, and evaluator caching.

Validation:

• Full suite: 3,562 passed, 10 skipped, 4 xpassed; 91.09% overall coverage and 100% for the new span helpers.
• All pre-commit checks and the Bandit security scan passed.
• Built wheel verified in a clean Python 3.12 environment with base dependencies only; span imports and pure helper calls require no AWS credentials.
• Bug bash: 1,000 generated span batches matched the original helpers across five trace filters, with input non-mutation checks.

[github-actions]

✅ No Breaking Changes Detected

No public API breaking changes found in this PR.

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[jariy17]

CI note: both red checks here are pre-existing and unrelated to this change. I'll address them in a follow-up PR.

Test Python 3.11 / 3.12 / 3.13 — payments/integrations/langgraph/test_stage3.py::TestRetryDelay::test_zero_delay_skips_sleep fails with Expected 'sleep' to not have been called. Called 898 times. Calls: [call(1), call(1), ...].

Root cause is a cross-test interaction that exists on main:

1. ragas constructs its AnalyticsBatcher at import time, which starts a daemon thread looping on time.sleep(1). It is not gated by RAGAS_DO_NOT_TRACK, and it outlives the test that imported ragas.
2. Our tests patch sleep via a module's own time reference (e.g. patch("...middleware.time.sleep")). That attribute lives on the shared stdlib time module, so the patch is process-global.
3. When that daemon thread wakes inside a patched window, the mocked sleep returns instantly, its loop spins, and the mock accumulates hundreds of phantom sleep(1) calls — breaking call-count assertions in whichever test holds the patch.

That makes it timing-dependent, which matches 3.10 passing while 3.11–3.13 failed, and the same test passing on rerun in #675. Neither the failing test nor the patch pattern is touched by this PR.

Compat (evaluation) — the compat job installs pytest plugins only, so v1.23.1's tests_integ/evaluation/test_third_party_adapters.py errors at setup with ModuleNotFoundError: No module named 'ragas' / 'deepeval' (16 collection errors). The same failure occurs on unrelated branches, and the job is continue-on-error: true, so the workflow run itself reports success.