Overhaul ETL framework with advanced pipeline architecture

[bjaus]

Complete overhaul of the ETL package from a minimal 3-file Pipeline[S,T] implementation into a production-grade framework.

Core changes:

• 4 job types (ScanJob, FetchJob, SeekJob, PageJob) with specialized runners replacing the single generic pipeline
• 40+ optional interfaces discovered via type assertion at construction — error handling, dead-lettering, panic recovery, lifecycle/stage hooks, filtering, progress reporting
• Checkpointing support for resumable jobs with epoch-based cursor tracking
• Resilience layer (rate limit → circuit breaker → retry) configurable per-stage
• Graceful shutdown with drain timeout and failureSignal coordination

Subpackages:

• batch/ — 9 pull-based collector implementations (Simple, Sized, Grouped, Striped, Windowed, Dedupe, Coalescing, Priority, Sorted)
• compose/ — Embeddable composables for common patterns (ErrorSkip, ErrorSkipN, ErrorStage, PanicSkip, Lifecycle, Progress) using *slog.Logger
• resilience/ — Retrier, TokenBucket/LeakyBucket rate limiters, RampingRateLimiter (6 curves), CircuitBreaker

Other:

• Internal dependencies replaced with stdlib (slog, sentinel errors)
• Documentation and examples use generic domain types throughout