fix(dlio): strip URI scheme from storage.storage_root (#392)

[FileSystemGuy]

Resolves #392.

Summary

When a user passes --params storage.storage_root=s3://bucket/path, every S3 upload fails because DLIO's obj_store_lib builds URLs with a doubled scheme — s3://s3://bucket/path/.... The reporter saw "nothing in S3 / no bucket created" as a result.

The doubled prefix originates in DLIO's obj_store_lib.ObjStoreLibStorage:

• storage_factory.py:49 passes args.storage_root straight in as namespace.
• obj_store_lib.py:159 stores it verbatim.
• obj_store_lib.py:399-401 unconditionally adds uri_scheme:// when constructing object URIs:

  if '://' in str(id):
      return id
  return f"{self.uri_scheme}://{self.namespace.name}/{id.lstrip('/')}"

So a storage_root like s3://mlperf-data/unet3d plus a key like data/unet3d/train/img_000_of_168.npz produces s3://s3://mlperf-data/unet3d/data/unet3d/train/img_000_of_168.npz. MinIO rejects every put, and Keith's command exits without writing anything.

--object mode happens to avoid this because _apply_object_storage_params sets storage_root = $BUCKET (bare bucket, no scheme). Users who hand-roll --params storage.storage_root=... hit it.

What this PR does

• Adds DLIOBenchmark._strip_uri_scheme and normalizes storage.storage_root in process_dlio_params so any <scheme>:// prefix is stripped before the override is handed off to DLIO. The bare-bucket form is untouched.
• Adds parametrized tests covering s3://, s3://bucket/, az://, gs://, and bare-bucket cases, plus the reporter's exact command shape.

The proper fix still belongs in DLIO's obj_store_lib — once the pinned fork is patched and pyproject.toml:127 is bumped, this normalization becomes a harmless no-op, since the normalized input is also valid input.

Test plan

• uv run pytest tests/unit -q → 1372 passed, 4 skipped.
• Manual: rerun the reporter's mlpstorage training datagen --model unet3d --num-processes 1 --params storage.storage_library=s3dlio --params storage.storage_type=s3 --params storage.storage_root=s3://mlperf-data/unet3d against MinIO and confirm objects land at s3://mlperf-data/unet3d/data/unet3d/train/img_*.npz (single s3://).

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[russfellows]

Approving per Curtis.