fix: replace ast.literal_eval with json.loads for eval CLI parsing [AE-1071]

[mjnovice]

Summary

• Replaced ast.literal_eval with json.loads in the LiteralOption CLI option class used by --input-overrides and --eval-ids
• ast.literal_eval cannot parse JSON values like null, true, false — causing eval runs with file input overrides containing these values to crash
• Added 13 regression tests covering JSON parsing with null, true, false, nested structures, and error cases

Root Cause

Part of the fix for AE-1071 — eval sets with file inputs crash because:

1. Backend serializes input overrides as JSON (with null/true/false)
2. Python CLI's LiteralOption uses ast.literal_eval which only understands Python literals (None/True/False)
3. Parsing fails → entire eval batch crashes

Test plan

• All 1591 existing tests pass
• New test_literal_option.py with 13 test cases covering JSON null/true/false parsing
• Lint passes (ruff check)

🤖 Generated with Claude Code

[mjnovice]

not needed anymore