perf(data): implement automatic field naming convention detection with LRU caching

[pinglin]

Because

• External packages like genai use camelCase json tags while Instill's schema uses kebab-case, causing unmarshaling incompatibility
• Nested Go struct fields using different naming conventions can cause silent data loss - when top-level structs use instill tags but nested external types use json tags with different conventions, field mappings fail and data is silently dropped
• Manual specification of naming conventions creates confusion and maintenance overhead for users
• Repeated field name resolution operations impact performance in high-throughput scenarios
• Code duplication existed between marshaling and unmarshaling logic for handling different naming conventions

This commit

• Implements automatic naming convention detection that tries multiple formats (kebab-case, camelCase, snake_case, PascalCase) and uses the one that exists in input data
• Prevents silent data loss in nested structures by recursively applying naming convention detection to all struct levels, ensuring external package types with camelCase/PascalCase/snake_case json tags are properly mapped even when the schema provides kebab-case field names
• Adds LRU cache with 200-entry capacity for struct field name mappings to improve performance
• Adds comprehensive benchmark tests showing 46% performance improvement for marshaling operations
• Maintains full backward compatibility while eliminating user configuration complexity

Performance Impact:

• Marshaling: 46% faster (801ns vs 1491ns), 53% less memory, 58% fewer allocations
• Unmarshaling: Minimal overhead (~2%) with automatic detection
• Memory bounded LRU cache prevents unlimited growth in long-running services