perf(rest): fetch remote scan plan tasks concurrently

[fallintoplace]

Summary

• Fetch each remote plan-task frontier with up to 8 concurrent REST requests.
• Keep breadth-first response order by storing each response at its handle index.
• Keep the existing once-only handle expansion and cycle protection.
• Add ordering and concurrency tests plus a fake REST latency benchmark.

Benchmark

Command:

go test ./catalog/rest -run '^$' -bench '^BenchmarkCollectScanTasks64Handles$' -benchtime=1x -count=5

Apple M1 Pro, darwin/arm64, 64 plan-task handles, and 10 ms simulated server latency per handle:

Fetch mode	Result
Serial baseline (limit 1)	733 to 777 ms/op
8 workers	90 to 93 ms/op

This is about an 8x reduction in the remote fetch portion. The 8-worker result is close to the expected 80 ms network floor, with HTTP and scheduling overhead.

Tests

• go test ./...
• go test -race ./catalog/rest -count=1
• go vet ./catalog/rest
• golangci-lint run --timeout=10m