Make ImePerfTest#testShowImeCold more accurate

In case killBaselineIme() will result the system busy on doing kill
process stuff and easier happen the locking contention in WM/ATM side
when measuring show IME with cold-lunch activity performance,

Make sure to start atrace after kill IME process being settled down
and dumping the trace for each iteration.

With this CL, following metrics can more accurate in reality without
being affected by unexptected delay:
- testShowImeCold_cv: 70 -> 10-20
- startInputOrWindowGainedFocus_mean: 6-8ms -> 2.8-4ms

Bug: 266708619
Test: atest ImePerfTest#testShowImeCold
Change-Id: Ibd004a36759e557c90bd6f156af3729d39f7721f