Introduce fair parallelism for trace part merge to prevent long-merge starvation

[hanahmily]

Problem

Long-running trace merges can delay shorter merge tasks and degrade compaction responsiveness.

Current context

• Trace merge loop currently uses global concurrency gating via mergeMaxConcurrencyCh in banyand/trace/merger.go.
• Merge and sidx merge latency are tracked but scheduling is not fairness-aware.

Proposal

• Add fairness-oriented merge scheduling (e.g., small/large queue classes or age-based prioritization).
• Ensure short merges are not indefinitely blocked by large merge jobs.
• Add observability for queue wait time and merge scheduling decisions.

Acceptance criteria

• Merge scheduling policy is deterministic and test-covered.
• Small merges complete within bounded delay under mixed workloads.
• Metrics/logs expose queue time, execution time, and scheduling class.