[cpu][windows]: compute cpu-total times from integer ticks

[skartikey]

Fixes #2110.

Times(percpu=false) converts each FILETIME half of the GetSystemTimes counters to float64 separately and then computes system = kernel - idle in floating point. On machines with a few days of uptime the cumulative counters are large enough that the conversion carries a rounding error of around 2^-30 s, and the error changes from call to call. As a result the returned System, Idle and User values are not monotonic even though the underlying Windows counters are, and consumers computing usage deltas occasionally see small negative values (see influxdata/telegraf#18141).

This change reassembles the full 64-bit tick counts from the FILETIME halves, does the subtraction on integers, and converts to seconds once, the same way perCPUTimes already does it. Since integer-to-float conversion and division by a constant are monotone, the reported values can no longer appear to decrease.

Running the simulation from #2110 against the new conversion shows 0 phantom decreases in 100000 intervals, versus 27674 with the current code.

[shirou]

Thanks for the detailed issue report and the reproducer in #2110.

The fix looks correct to me:

• Reassembling the FILETIME halves and doing the subtraction on integers is exactly what perCPUTimes already does, so both paths are now consistent.
• Monotonicity holds by construction: integer-to-float64 conversion and division by a positive constant are both monotone non-decreasing, so the reported values can no longer appear to decrease while the underlying counters don't.
• kernel - idle is now exact, so an interval whose true delta is 0 reports exactly 0 instead of accumulated rounding noise.

kernel >= idle is guaranteed by the GetSystemTimes contract (kernel time includes idletime) described in MSDN, so the subtraction is safe.

LGTM. Thanks!