v0.10.0

Paging a collection no longer costs a seek per skipped row

skip was applied by pulling and discarding, and each discarded row was a key lookup and a get — the fetch and decode of a document the caller had asked to pass over. The offset is now taken at the source and reached by stepping fjall's key iterator, which reads keys without the values behind them.

Measured over 20k rows of ~1KB, 400 to a page, against a full scan that is drained rather than counted:

	paged walk over the whole collection	vs one full scan
0.9.0	1.089s	18.6x
0.10.0	0.093s	1.6x

The push-down is declined where anything between the source and the page drops or reorders rows — a scanned filter, a blocking sort, an or plan. Those keep the pipeline skip, which is correct and merely pays for what it passes over.

Added

• NitriteMapProvider::skip_keys_from_start, defaulted to None so every backend outside this repository keeps working unchanged. The fjall implementation declines while a transaction is in scope, because the keys visible there include the scope's uncommitted inserts and exclude its tombstones, and a count over the committed partition alone would land on the wrong row.
• EntryIteratorProvider::skip_entries, same shape, for providers that iterate entries.

Correction

The figures published in #24 — "68.8x to 40.4x" — were measured against find(all()).count(), and an unfiltered count() is answered from the map size without reading a single document. The absolute times were right; the denominator was not. Read as 18.6x to 9.8x, with this release taking it to 1.6x.

Full changelog: https://github.com/nitrite/nitrite-rust/blob/main/CHANGELOG.md