Microbenchmark - Fix precision loss in kernel launch overhead benchmark

[alephpiece]

Description

The original code uses integer division in the time calculation: (end_tv.tv_usec) / 1000. Since an empty kernel launch overhead is typically very small (often less than 1ms), this integer division truncates the microsecond component to 0, leading to significant precision loss and incorrect results.

Major Revision

• Changed the divisor from 1000 to 1000. (floating-point literal). This forces floating-point arithmetic, preserving the microsecond precision required for accurate overhead measurement.

[guoshzhao]

Thanks @alephpiece for pointing out this issue. We will verify and review this PR.