- Currently we always compute the vector index size, irrespective of
whether this size was cached in the snapshot's `cachedMeta` or not. We
use the cached value when triggering the searcher start callback, but do
not check if it exists in the `cachedMeta` before calling the `Size()`
API.
- Due to recent changes, the output of the `Size()` API was fixed but at
the cost of additional computation to get a precise estimate. This
results in a perf regression due to the aforementioned point.
- Fix this issue by performing a `LoadOrStore` operation for the size
instead, where we perform a read to check if the size was cached, and
only compute it if not cached.
- Refactor `cachedMeta` to use a `sync.Map` instead of a map with a RW
mutex as its more suited for the `write-once-read-many` style of usage
which we use it for in this perspective.
- Fix the `TotKNNSearches` to be equal to the number of KNN searchers
created, which mirrors the FTS equivalent `TotTermSearchersFinished`,
currently this stat is equal to `N` * `X` where `N` is the number of
segments, refactor it to report just `X`.