feat: add post-loader data contract validation for SFT/DPO/Online RL/Agentic

[Hpward-Ling]

What does this PR do?

Adds a post-loader data contract validator that checks required fields, types,
list lengths, and nested messages for SFT, DPO, Online RL, and agentic inputs
after the user dataset loader runs, before expensive model/worker
initialization. Errors are aggregated into a bounded report (instead of stopping
at the first), with sample index, field path, expected/actual type, and a
concrete correction hint per error.

Three entry points:

1. Python API — from areno.api import validate_contract → returns a
   ContractReport with .ok, .errors, .warnings, .total_scanned.
2. CLI standalone — areno data inspect --dataset-path ... --contract --mode {sft,dpo,online_rl,agentic}
   with --json for CI-friendly structured output.
3. CLI integrated — areno train ... --validate-data-contract runs the
   same validator after dataset loading, before model init. Default is
   --no-validate-data-contract (backward compatible).

Related issue

Fixes #214

Type of change

• ✨ New feature

How was it tested?

All tests are CPU-only, no GPU or external services required.

# Core validator unit tests (all modes, boundary, malformed, disabled/default)
pytest tests/test_data_contract_cpu.py -v

# Data inspect CLI tests
pytest tests/test_data_inspect_cli_cpu.py -v

# Train CLI integration tests
pytest tests/test_train_cli_contract_integration_cpu.py -v

# Run all three together
pytest tests/test_data_contract_cpu.py tests/test_data_inspect_cli_cpu.py tests/test_train_cli_contract_integration_cpu.py -v