fast, reliable time stamp counter for ARM

```
/* There is no simple equivalent to x86's rdtsc on ARM.
* There is the Cycle CouNT (CNT) register, but it requires kernel support
* to make it accessible from user mode, and it can be reset, resulting in
* potential conflicts with the app. Others seem to map /dev/mem and figure
* out where the hardware counter is but this does not seem portable w/o
* a kernel driver.
* For now we punt on having kstats on by default and live with an expensive
* system call.
*/
return query_time_micros();
```

The issue only points to query_time_micros() and the ARM timing limitation; start by tracing that call and reviewing how ARM user-mode counters, /dev/mem, kernel support, and kstats are handled. Done means an agreed portable, fast ARM timestamp-counter path replaces the expensive syscall without reset conflicts.

Written by the indexing model from the issue text.