[Feature] datasets fs: inspect and download files under a task (ls/get/download)

[xdlkc]

Feature Category

• Sandbox
• Actions
• Deployments
• SDK & API
• Envhub
• CLI
• Performance & Optimization
• Documentation & Examples

Problem Statement

There is currently no way to inspect or fetch individual files inside a single
task from the OSS dataset registry. Users can list task IDs (datasets tasks)
but to look at or download a task's contents they must reach for raw ossutil
and reconstruct the datasets/{org}/{dataset}/{split}/{task}/ key layout by
hand. Two concrete gaps:

1. No file listing per task: no command answers "what files does task X
   contain?"
2. No file read / download: no command prints a single task file to stdout
   or downloads a task file/directory to a local path.

Proposed Solution

Add a datasets fs subcommand group (alias files) with three operations,
backed by new registry/client methods.

1. Registry layer (OssDatasetRegistry / BaseDatasetRegistry /
   DatasetClient):

   • list_task_files(org, dataset, split, task_id, path="") -> list[TaskFile]
     — list files under a task, paths relative to the task root.
   • get_task_file(org, dataset, split, task_id, path) -> bytes | None
     — read one task file by relative path; None when the object is absent.
   • New TaskFile dataclass (path: str, size: int | None).
   • Support both layouts: directory-style tasks ({task}/...) and
     single-file tasks ({task}.json directly under the split).

2. CLI layer (datasets fs ...):

   • datasets fs ls --org --dataset --split --task [--path] — list files.
   • datasets fs get --org --dataset --split --task [--path] — print one
     file to stdout (resolves the only file automatically when the task has
     exactly one).
   • datasets fs download --org --dataset --split --task --path --dest —
     download a single file or a whole directory subtree to a local path.
   • JSON output (-o json) for ls/get/download.

3. Path safety: relative task paths are normalized and validated — reject
   absolute paths and .. traversal, drop ./empty segments — so a crafted
   --path cannot escape the task root.

Notes

• Split out from [Feature] datasets tasks: server-side pagination, continuation token caching, and --filter support #1063 / PR perf(datasets): cache OSS bucket, add server-side pagination and --filter for tasks #1064: [Feature] datasets tasks: server-side pagination, continuation token caching, and --filter support #1063 covers the datasets tasks
  pagination/cache/--filter work; this issue tracks the independent
  datasets fs file-access feature.
• The CLI command module (rock/cli/command/datasets.py) will be refactored
  toward a clearer object-oriented structure as part of this change (per review
  feedback on PR perf(datasets): cache OSS bucket, add server-side pagination and --filter for tasks #1064): path normalization and output handling extracted into
  dedicated TaskPath / OutputWriter helpers, with the module-level helper
  functions folded into the command class.