performance during repeated binfuse::sharded_filter::add_shard()

[oschonrock]

the unmapping and re-mmapping take ~400ms for 256 shards. Could be mitigated with a "predictive growth plus trim at the end" strategy.

This may seem like a lot, however for real data, unlike in a test, where building the whole filter (for hibp for example) takes 2-3minutes for 1billion entries, the 400ms is not significant.

The primary use case for sharded_filter is for such large datasets, so this would probably be a microptimisation.