Add additional dataset format support (pyarrow, csv, tsv, etc.)

[Cameron7195]

Summary
Support viewing non-JSONL datasets (PyArrow-based, CSV, TSV) in a way that still gives users a good inspection experience.

Problem
Many real-world ML datasets aren’t stored as JSONL; they may be in PyArrow/Parquet, CSV, or TSV. Right now, the viewer is effectively JSONL-centric, which limits its usefulness for ML engineers working with diverse formats.

Proposed Solution

• Define a minimal abstraction for “row-based dataset” independent of underlying storage.
• For each supported format:
  • CSV/TSV:
    • Parse headers as column names.
    • Treat each row as a flat object; if a cell contains JSON, optionally detect and pretty-print it.
  • PyArrow / Arrow-backed formats (initially optimistic / read-only assumptions):
    • Use a Node/JS-accessible reader or a conversion step to present rows and columns.
    • Preserve column names and basic types; optionally detect JSON strings similarly.
• Keep the UI consistent: same pretty-printing and JQ-style key selection concepts where applicable.

Acceptance Criteria

• User can open CSV/TSV files and:
  • See rows/columns with column headers.
  • Optionally expand JSON-like cells as nested JSON views.
• Basic support for at least one Arrow-based dataset path (even if via a conversion step or a restricted subset).
• Errors for unsupported or malformed files are clear and non-crashing.