test(integ): absorb iam:SimulatePrincipalPolicy throttling across integ suites

[jam-jee]

Issue

Integration tests run under pytest -n auto (dozens of xdist workers). Many of them resolve/validate an IAM execution role via resolve_and_validate_role — invoked from TrainDefaults.get_role, ModelBuilder.build()/.deploy(), Pipeline create/upsert, and the feature-processor scheduler. That function calls the low-TPS iam:SimulatePrincipalPolicy API (plus a get_role for the trust check). With many workers hitting it at once, IAM throttles the request:

botocore.exceptions.ClientError: An error occurred (Throttling) when calling the
SimulatePrincipalPolicy operation (reached max retries: 4): Rate exceeded

_evaluate_permissions only degrades AccessDenied/NoSuchEntity; Throttling re-raises and fails the test — often during fixture setup, not just the call phase.

Fix (test layer only — no SDK source change)

Add an identical, self-contained mitigation to the serve, train, and mlops integ conftests:

1. _configure_boto_adaptive_retries (autouse, session-scoped) — sets adaptive retries via the AWS_RETRY_MODE / AWS_MAX_ATTEMPTS environment variables. Env vars apply to every boto3 client created in the worker, so the IAM calls ride out transient throttling whether the resolver falls back to the default session or builds its client from an explicitly-passed Session. Pre-existing env values are restored on teardown.
2. pytest_runtest_makereport — belt-and-suspenders: converts a residual SimulatePrincipalPolicy throttle that survives retries into a skip (checked in both setup and call phases). Throttling on any other operation still fails loudly.

Why not the previous serve-only approach?

The prior serve conftest set a retry Config on boto3.DEFAULT_SESSION. That only helps when the resolver falls back to the default session — several serve tests (and the train/mlops fixtures) pass their own Session(boto_session=...), whose IAM client carried botocore's default 4-attempt policy and was unaffected. Env vars close that gap and let all three suites share one implementation.

Scope

• New parent sagemaker-train/tests/integ/conftest.py covers train/, ai_registry/, and jumpstart/.
• sagemaker-core integ is intentionally not modified — resolve_and_validate_role has no call path there, so it never hits SimulatePrincipalPolicy.

Testing

• All three conftests compile (py_compile) and lint clean.
• Unit-checked the throttle-detection helper: matches Throttling/ThrottlingException/RequestLimitExceeded only on SimulatePrincipalPolicy; ignores other ops, other error codes, and non-ClientError exceptions.

Type of change

• Bug fix / test-infrastructure fix (non-breaking; test-layer only)

Checklist

• I have read the CONTRIBUTING doc
• Changes are limited to test conftests; no SDK source or public API is touched
• No new dependencies introduced

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[mujtaba1747]

I think retries are a good idea.

But instead of skipping can we do the following:
Can we go to IAM team to increase rate limit
Or add feature to skip iam validations

[lucasjia-aws]

with batch-call we should not actually go into skip branch right? But we should keep monitoring if tests skips happen a lot with this change