architecture: refactor to 3-stage pipeline with separated Reader, Decoder, Encoder goroutines

[more-shubham]

Currently processImage does file I/O, decode, and encode sequentially in one function inside a single worker goroutine. Separating into 3 pipeline stages allows each stage to run concurrently and makes bottleneck identification precise.

Scope:

• internal/pipeline/reader.go — file open and raw bytes read, outputs to decode channel
• internal/pipeline/decoder.go — JPEG/PNG decode from raw bytes, outputs image.Image to encode channel
• internal/pipeline/encoder.go — WebP encode and file write, outputs Result
• Each stage has its own bounded channel and worker count
• Add stage-level metrics: time spent in read vs decode vs encode separately

Architecture:

JobsChan → [Reader goroutines] → rawChan → [Decoder goroutines] → imgChan → [Encoder goroutines] → ResultsChan

Acceptance Criteria:

• Each stage independently configurable in worker count
• Stage timing visible in summary.json under stage_metrics
• Memory profile shows decode buffer reuse across pipeline
• All existing tests pass against new architecture
• BENCHMARKS.md updated with stage-level timing breakdown