…upport

The frontend was calling the HF datasets-server API directly from the
browser to fetch dataset split/subset metadata. This does not work for
private datasets because the external datasets-server API does not
reliably serve metadata for private repos even with Bearer tokens.

The dataset preview (check-format) already worked for private datasets
because it routes through the backend which uses the `datasets` library
with proper token support.

Fix: Add a new POST /api/datasets/splits backend endpoint that uses
`get_dataset_config_names` and `get_dataset_split_names` from the
`datasets` library with the HF token. Update the frontend hook to call
this backend endpoint via authFetch instead of the external
datasets-server.

Fixes unslothai#4962

for more information, see https://pre-commit.ci

- Catch HfHubHTTPError and propagate the original HTTP status code
  (404, 403, etc.) instead of always returning 500
- Raise HTTPException when all per-config split lookups fail, so the
  UI receives an actionable error instead of a misleading empty-success
- Re-raise HTTPException without wrapping to avoid double-wrapping

When get_dataset_split_names fails with HfHubHTTPError for every
config, extract and propagate the original HTTP status code (401, 403,
etc.) instead of converting it to a generic 500. This ensures the
frontend receives actionable status semantics for auth errors.

- Move HfHubHTTPError import outside try block so the outer except
  clause cannot raise NameError if the import fails
- Catch DatasetNotFoundError explicitly and return 404 instead of 500
  for nonexistent/inaccessible datasets
- Remap upstream HF 401 to 403 to prevent authFetch from misinterpreting
  HF auth failures as studio session expiries (which triggers logout)
- Reset last_config_status to 500 in the generic exception handler to
  avoid mismatched status/message when mixing exception types
- Validate dataset_name with min_length=1 to reject empty strings early

for more information, see https://pre-commit.ci

The previous regex rejected valid HF repo IDs with leading underscores
(e.g. org/_dataset). Delegating to HF's own validator ensures we match
their exact rules: alphanumeric, -, _, . with no -- or .. sequences.

for more information, see https://pre-commit.ci

Remove autouse from conftest.py fixtures and declare fake_datasets
and fake_structlog explicitly in each test function that needs them.
This prevents the stubs from affecting unrelated tests in the repo.

for more information, see https://pre-commit.ci

The previous commit accidentally duplicated fake_datasets, fake_structlog
in test function parameters due to sequential sed commands matching
already-modified lines. This corrects all 7 affected test files.

Tests that import fastapi or studio backend modules now call
pytest.importorskip("fastapi") at module level so non-studio CI
pipelines skip them gracefully instead of failing at collection.

The previous import from huggingface_hub.utils._validators is a private
module that could break on dependency upgrades. Use the public path instead.