[azure-ai-ml] Fix job download and streaming regressions

[Kshitij Chawla (kshitij-microsoft)]

Description

Fixes two azure-ai-ml 1.35.0 regressions introduced when job and datastore consumers moved to the shared TypeSpec arm_ml_service client.

Named output download

RunHistory reports output asset types in PascalCase (UriFolder, UriFile, MLTable), while the shared ARM client exposes snake_case enum values (uri_folder, uri_file, mltable). The exact comparison filtered out valid named outputs, causing MLClient.jobs.download(..., output_name=...) to complete without downloading anything.

Asset types are now normalized case- and separator-insensitively before comparison.

Fixes #48941.

Identity-based log streaming

The same enum value change activated the direct-datastore log path for URI-folder outputs. That path generates a short-lived SAS and requires a string account key or SAS token. Identity-based datastores instead provide a TokenCredential, which caused SAS generation to raise:

TypeError: argument should be a bytes-like object or ASCII string, not 'ChainedTokenCredential'

Streaming now falls back to the service-provided RunHistory log URLs when the datastore credential cannot sign a SAS. Key- and SAS-based datastores continue using the direct path.

Testing

• python -m pytest tests/job_common/unittests/test_job_ops_helper.py -q -p no:warnings
• Result: 6 passed, 15 skipped
• Added regression coverage for PascalCase and snake_case output types.
• Added regression coverage for string, TokenCredential, and missing datastore credentials.

Notes

The changes are limited to hand-written azure-ai-ml operations code, tests, and changelog. No generated code or service wire contract is changed.

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).
10 pipeline(s) were filtered out due to trigger conditions.
There may be pipelines that require an authorized user to comment /azp run to run.