[BOO] Sync after profiling to fix flaky event loss

[rkayaith]

Synchronize GPU in the BOO driver's profiling loop. We've been getting flaky failures in tests/kernel/boo/driver/cli_test.py with dispatches not being captured by the profiler, and I found that synchronizing in two places resolves the issue:

1. Before each prof.step() — ensures in-flight GPU kernels complete before the profiler advances state
2. After the profiler context exits — ensures profiler cleanup is complete before the next profiling session starts

The torch profiler context already includes a synchronize in its __exit__ so that all events are captured, so this seems to be something to do with needing to synchronize after the profiler's cleanup.

Fixes #1327