feat(benchmarks): Add log and download source files

[dolaameng]

feat(benchmarks): Add publish and log commands, support dataset push, and improve download

TODO: release kaggle-sdk-python new version and use it here.

Publish command (new)

Adds kaggle b t publish <task> [--publish-backing-notebook] to make a benchmark task public.

• Handles partial states: if a task is already public but the backing notebook is private (and the flag is set), it will publish only the notebook.
• Returns clear warnings if no backing notebook is associated with the task.

Log command (new)

Adds kaggle b t log <task> [-m <model> ...] to view execution logs.

• Streaming: Active (RUNNING) runs stream live output via SSE; completed/errored runs print the persisted log.
• Robustness: Gracefully handles QUEUED runs (which have no logs yet) by catching the server's 404/400 per-run, printing a friendly indicator, and continuing to show logs for other runs instead of crashing.
• Multi-model: Logs for each model are printed sequentially (not interleaved), with a header showing run ID and state, a line-count footer, and a final summary.

Push with dataset (new)

Adds -d / --kaggle-dataset to kaggle b t push to attach Kaggle datasets as data sources.

• Visual warning if re-pushing a task without -d would detach previously associated datasets.
• Clean validation checks to warn about unresolved dataset slugs on the server.

Download improvements

• --include-source / -s: Also download the kernel session's source notebooks alongside output files.
• --force / -f: Re-download and overwrite previously downloaded runs instead of skipping them. Downloads to a staging directory first so the previous output is preserved if the download fails.
• Summary line: Prints a final Done: X downloaded, Y skipped. count (and outputs Done: 0 downloaded. consistently on early returns).

Refactors & Code Quality

• TTY-Aware Colors: Replaced all raw ANSI escape codes (\033[) with static utility helpers (_bold(), _warn(), _error()) that automatically strip color codes when the output stream is redirected (non-TTY, pipes, logs, agents).
• Format Improvements:
  • Unified _format_model_hint() to fix model name quoting inconsistencies.
  • _print_log_entry() now prints fallback dictionaries with formatting via json.dumps() instead of Python's raw repr().