[fuzz] Handle statistical error in normalization of CPU times

Since the normalization values are experimentally determined they have
potential for error. Previous experience suggests to me that this error
might be significant, so adding support for quantifying and propagating
it in the calculations.

Specifically multiple iterations of normalization microbenchmarking are
now run to produce a mean and error which is then propagated through to
later calculations.

Since normalization is now considered an experimental value, it is no
longer used as a substitute for burn-in, so the overall experimental
harness is reworked to be more readable/sequential. If there are any
normalization or experimental tasks to be run, then burn-in (if enabled)
is run before hand, so that thermal throttling/etc is minimized.

The terminology around normalization has been, well normalized, since it
was confusingly referred to as perf score and benchmarking in different
parts of the code.

Also re-shuffled members in structs and documentation to be better
grouped together, since the burn-in values and normalization ones were
interleaved in places.

Bug: 524444910
Change-Id: Ic7b31173b2aaa5950f46fd789284ec38ab16f7b9
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/339415
Reviewed-by: dan sinclair <dsinclair@chromium.org>
Commit-Queue: Ryan Harrison <rharrison@chromium.org>