Exclude input plan's execution time from `elapsed_compute` statistic

**Is your feature request related to a problem or challenge? Please describe what you are trying to do.**

Baseline metrics have an `elapsed_compute` timer that counts the CPU time of one physical plan. It should not include the time that polling from inputs.

**Describe the solution you'd like**

I find these plans have the issue:
- `CoalesceBatchesExec`
- `SortMergeJoinExec`

I'll move the timer after the input record batch is ready.

**Additional context**

An example I got:
```plaintext
Executor executed plan, request_id:7, plan_and_metrics: SortExec: [minute@0 ASC NULLS LAST], metrics=[output_rows=60, elapsed_compute=47.392µs, spill_count=0, spilled_bytes=0]
CoalescePartitionsExec, metrics=[output_rows=60, elapsed_compute=10.018µs, spill_count=0, spilled_bytes=0, mem_used=0]
ProjectionExec: expr=[time_bucket(cpu.timestamp,Utf8("PT60S"))@0 as minute, MAX(cpu.usage_user)@1 as max_usage_user, MAX(cpu.usage_system)@2 as max_usage_system, MAX(cpu.usage_idle)@3 as max_usage_idle, MAX(cpu.usage_nice)@4 as max_usage_nice, MAX(cpu.usage_iowait)@5 as max_usage_iowait], metrics=[output_rows=60, elapsed_compute=3.225µs, spill_count=0, spilled_bytes=0, mem_used=0]
AggregateExec: mode=FinalPartitioned, gby=[time_bucket(cpu.timestamp,Utf8("PT60S"))@0 as time_bucket(cpu.timestamp,Utf8("PT60S"))], aggr=[MAX(cpu.usage_user), MAX(cpu.usage_system), MAX(cpu.usage_idle), MAX(cpu.usage_nice), MAX(cpu.usage_iowait)], metrics=[output_rows=60, elapsed_compute=172.435µs, spill_count=0, spilled_bytes=0, mem_used=0]
CoalesceBatchesExec: target_batch_size=4096, metrics=[output_rows=60, elapsed_compute=91.765µs, spill_count=0, spilled_bytes=0, mem_used=0]
RepartitionExec: partitioning=Hash([Column { name: "time_bucket(cpu.timestamp,Utf8(\"PT60S\"))", index: 0 }], 6), metrics=[repart_time{inputPartition=5}=33.112µs, send_time{inputPartition=5}=26.289µs, fetch_time{inputPartition=5}=147.221616ms]
AggregateExec: mode=Partial, gby=[time_bucket(timestamp@0, PT60S) as time_bucket(cpu.timestamp,Utf8("PT60S"))], aggr=[MAX(cpu.usage_user), MAX(cpu.usage_system), MAX(cpu.usage_idle), MAX(cpu.usage_nice), MAX(cpu.usage_iowait)], metrics=[output_rows=60, elapsed_compute=334.918µs, spill_count=0, spilled_bytes=0, mem_used=0]
CoalesceBatchesExec: target_batch_size=4096, metrics=[output_rows=480, elapsed_compute=73.715112ms, spill_count=0, spilled_bytes=0, mem_used=0]
FilterExec: hostname@1 = host_804 OR hostname@1 = host_3829 OR hostname@1 = host_2786 OR hostname@1 = host_3242 OR hostname@1 = host_2509 OR hostname@1 = host_648 OR hostname@1 = host_204 OR hostname@1 = host_1769 AND timestamp@0 >= 1662349096000 AND timestamp@0 < 1662352696000, metrics=[output_rows=480, elapsed_compute=70.708739ms, spill_count=0, spilled_bytes=0, mem_used=0]
ScanTable: table=cpu, parallelism=8, order=None, , metrics=[]
```
Here `CoalesceBatchesExec` takes ~70ms. It includes the child node's execution time.

Start by locating the implementations of CoalesceBatchesExec and SortMergeJoinExec and the elapsed_compute metric timing. Check how input polling is included in each plan's timing, then verify that completed metrics exclude time spent waiting for input while still measuring execution work.

Written by the indexing model from the issue text.