feat: pluggable components, IR layer, multi-backend, REST API server

[rafafrdz]

Summary

Pluggable Component System (#112, #113)

Inspired by datastream-connector-engine:

• Trait interfaces: TeckelReader, TeckelTransformer, TeckelWriter with standard read/transform/write methods
• PluginRegistry: Register/resolve components by name or class path, dynamic class loading via reflection
• ComponentResolver: Core namespace + full-class-path resolution (like datastream-connector-engine's components.py)
• Custom transformation: custom: in YAML delegates to a registered TeckelTransformer plugin
• YAML components: section: Configure readers/transformers/writers by alias with core name or external class

IR Layer (#102)

• IRNode sealed trait with Scan, Project, Filter, Aggregate, Sort, JoinNode, Sink, Transform
• IRPlan DAG with topological ordering, fan-out analysis, root/sink queries
• IRBuilder converts existing Context[Asset] to IRPlan
• Optimizer with predicate pushdown pass

Multi-Backend Abstraction (#103)

• Backend trait with execute and dryRun methods
• SparkBackend implementation (bridges to existing semantic layer)
• DataFusionBackend placeholder for future integration
• config.backend field in YAML

REST API Server (#105)

• Lightweight embedded server using JDK com.sun.net.httpserver (zero extra deps)
• Endpoints: health, dry-run, doc generation, graph visualization, validation
• CLI: --server flag with optional --port N

Closes #102, #103, #105, #112, #113

Test plan

• sbt clean compile passes
• CI green