Processor durability and processor parallelism

[novatechflow]

Summary

This implements #113 by removing an unnecessary .index download from the processor decode path instead of parallelizing two S3 reads where only one was actually used.

It also fixes a PITR bug by making restore truncate the final segment at the requested cutoff and regenerate a matching .index, rather than only restoring whole segments.

While tracing that path, I also tightened PITR/restore behavior in storage so the final recovered segment can be truncated at the requested cutoff and a matching .index is regenerated.

Changes

• stop downloading .index in the SQL and Iceberg processor decoders
• keep .kfs/.index pairing validation in discovery
• add regression coverage for the decoder behavior
• add exact segment truncation for PITR/restore
• rebuild the final restored .index after truncation
• add recovery and CLI regression coverage with real batch fixtures

Testing

• go test ./pkg/storage ./cmd/kafscale-cli -count=1 -timeout=120s
• go test ./... -count=1 -timeout=120s in addons/processors/sql-processor
• go test ./... -count=1 -timeout=120s in addons/processors/iceberg-processor

Checklist

• Added/updated unit tests for new logic
• Added/updated e2e coverage for bug fixes
• Added license headers to new files