Fix/issue214 Validate post-loader data contracts by training mode

[pzx-dny]

What does this PR do?

Add post-loader data contract validation that checks dataset schema before expensive model/worker initialization. Supports SFT, DPO, online RL (GSPO/GRPO/PPO), and agentic modes with bounded error collection and human-readable/JSON output formatting.

Key changes:

• New areno/api/data_validation.py — core validation module (stdlib only, no new dependencies)
• Integrate validation into areno/cli/train.py:run() after dataset loading, before build_trainer()
• New tests/test_data_validation_cpu.py — 26 CPU unit tests

Behavior: Valid data passes silently. Invalid data prints aggregated errors (sample index, field path, expected/actual type, fix hint) to stderr and raises click.UsageError to block training. No new CLI flags or subcommands.

Related issue

issue#214

How was it tested?

$ pytest tests/test_data_validation_cpu.py -v
============================== 26 passed in 0.04s ==============================

Test coverage:

• SFT mode (7): valid data, missing field, empty value, wrong type, mixed errors, valid_rows count
• DPO mode (5): prompt/response style, chat message list style, mismatched types, missing field, None value
• Online RL mode (4): GSPO/GRPO/PPO valid, non-dict row detection
• Agentic mode (1): auto-detection via agent_fn
• Boundary (3): max_errors limit, empty dataset, unknown algo
• Output format (6): human-readable and JSON output assertions

Hardware limitation: macOS without GPU — tests load the module directly to bypass the areno.api import chain (which requires torch). On Linux GPU environments, from areno.api.data_validation import ... works directly.

Checklist

• The PR title summarizes the contribution.
• Linked the related issue in the description (if any).
• Existing tests pass (pytest tests/ -k cpu).
• New behavior is covered by tests.
• Described the test commands run and any hardware limitations.
• Public API / CLI changes are additive and backward-compatible (see CONTRIBUTING.md).

Breaking change details

No breaking changes. Validation runs by default but is completely silent for valid data — no impact on existing behavior.