allow scopes to flush between each sample.

[tgolsson]

Benchmarks run fine to the end, but then dies to memory kill at exit. The scope at the top seems to prevent flushing, so all data is kept in memory until we exit. By only delaying flushing until after each benchmark we can still hide it from measurement (assuming bench_function isn't measured) but prevent memory growth between benches. Adding the new_frame just for good measure as it allows dropping the data progressively too.

[tgolsson]

@andreiltd Did you have anything that you were investigating beyond this that you'd want to combine this with? Did you find a good explanation for why this improved perf beyond me guessing? :D

[andreiltd]

LGTM, although I'm not sure if collecting frames affects the profiled function in any way or if this occurs entirely outside the measured scope?

[tgolsson]

From the criterion docs, my understand is that only b.iter() is measured. So it should be unchanged. Which makes your detected diff a even weirder.