Improve Top SQL performance in sysbench

## Enhancement

After Top SQL is introduced sysbench workload degrades from 6% to 10% in different workloads when CPU utilization is full. This issue is to track the optimizing process.

Possible optimizations:

**AttachSQLInfo**:

- [ ] Reduce allocations in `linkSQLTextWithDigest` and `linkPlanTextWithDigest` #33378

- [ ] Reduce number of calls to `AttachSQLInfo`

For 10 QPS oltp_point_select, 30 QPS calls to this function can be observed. We should be able to reduce the calls to 20 QPS (10 for SQL digest, 10 for Plan digest).

- [ ] Further reduce number of calls to `AttachSQLInfo`

It is possible to only call `pprof.SetGoroutineLabels` once per request, as long as we attach a ID to the goroutine, instead of attaching SQL digests and plan digests. This also fully resolves the problem of "Unknown Plan".

- [ ] Hack

`pprof.SetGoroutineLabels` builds the `g.labels` via `runtime_setProfLabel(unsafe.Pointer(ctxLabels))`. So may be we can build and maintain our own map so that it can be reused instead of allocated each time.