include/arch_interface.h: Clarify platform-dependencies

Clarify a few restrictions that seem obvious but weren't specified:

+ On MP systems, counters are often asynchronous between cores, so you
  need to make sure all the values read are on the same CPU.

+ Many platforms have a 32 bit (or smaller, c.f. Cortex M SysTick)
  internal counter, but are being forced to return a 64 bit quantity
  that will behave badly on rollover (you can't just subtract two
  times anymore!).  Demand that the existing arch_timing_cycles_get()
  API (which didn't seem to have useful much to do) implement rollover
  correction on the subtraction.

This allows an Xtensa implementation based on direct CCOUNT reads on
intel_adsp to be correct per the API docs.

Signed-off-by: Andy Ross <andrew.j.ross@intel.com>