feat(data): add streaming JSONL and loader-output quality scanner

[Liu14159]

Summary

Add a streaming JSONL and loader-output quality scanner that inspects dataset files line-by-line without loading the full dataset into memory. Reports blank lines, JSON parse errors, non-object records, and schema surprises (missing required keys) with bounded, redacted previews of bad entries. Enables users to validate data quality before starting expensive training runs.

Closes #213.

What changed

• areno/api/data.py — scan_jsonl_stream (stream file/stdin line-by-line), scan_loader_output (scan custom loader records), ScanIssue/ScanReport dataclasses, format_scan_report (human-readable + JSON output), redaction (_redact_preview) and truncation (_PREVIEW_MAX_CHARS=200) helpers
• areno/cli/diagnostics.py — areno scan-dataset CLI command with --required-keys, --max-issues, --json, --loader-fn options, supporting file and stdin input; exit code 1 when issues found
• areno/cli/main.py — register scan-dataset in the command table
• tests/test_scan_dataset_cpu.py — 30 tests covering core scan logic, malformed input, boundary values, redaction, bounded memory (5000-line simulation), CLI paths (file/stdin/json/loader-fn/missing-file), and --help listing

Design

• Streaming, bounded memory: File is read line-by-line; memory does not scale with file size. max_issues caps stored ScanIssue entries; excess counted in truncated_issues.
• Issue categories: blank (whitespace-only line), json_error (JSON parse failure), non_object (valid JSON but not a dict), schema (dict missing required keys). Each issue records 1-based line number, detail string, and truncated/redacted raw preview.
• Redaction: _redact_preview masks values of common secret keys (api_key, token, secret, password, etc.) in previews so sensitive training data never appears in logs or CLI output.
• Defaults are backward compatible: required_keys=() (no schema check), max_issues=50. The command is purely additive — no existing CLI command, trainer, or config is modified.
• No new dependencies: Uses only stdlib (json, re). No external database, no sandbox, no model initialization.
• Loader-fn path: --loader-fn imports a user .py file, calls load_training_dataset(path, default_loader=...), and scans the returned records with scan_loader_output. Receives a no-op default_loader so scanning doesn't pull in the HF datasets stack.
• Output formats: format_scan_report(report, json_output=False) for human-readable summary + per-issue lines; json_output=True for structured JSON via ScanReport.to_dict().

Testing

30 CPU tests across 5 test classes:

• ScanJsonlStreamTest (13) — clean file, blank lines, JSON errors, non-object, schema check, error recovery, preview truncation, secret redaction, max_issues bounds (including 0 and negative), stdin input, source tracking
• ScanLoaderOutputTest (4) — clean records, non-dict, schema missing keys, source label
• FormatScanReportTest (3) — JSON output validity, human-readable status, issue line display
• ScanDatasetCliTest (9) — clean/bad file exit codes, required-keys, json output, stdin, max-issues, --help listing, loader-fn scan, loader-fn missing file
• BoundedMemoryTest (1) — 5000-line file with mixed content, verifies counts and issues cap