feat(scanner): plan _rowoffset and the DataFrame entry points

[wjones127]

Closes the last gap between the logical path and Scanner: _rowoffset, plus the DataFrame surface the logical path makes possible.

A row's offset is its position in the dataset once deletions are accounted for, so computing one needs every earlier fragment's row count and deletion vector. AddRowOffsetExec::try_new is the only physical node constructor in the read path that does I/O, which is exactly the shape this design exists to remove: the load moves into stage 2, and a rule hands the result to the node.

With that, the new path plans every query shape the imperative one can, so every scanner test compares the two directly rather than one deferring to the other.

One behavioural note on the frame surface: a row's _created_at_version / _last_updated_at_version come from its position within its fragment, so only an ordered scan can produce them and the take above a search cannot. A frame with no projection of its own asks for every column the leaf advertises, so the take drops those two rather than asking the read for something it cannot return.

LanceContextExt / LanceDataFrameExt fall out of the rest of the work rather than being built for it. The scanner exposes one fixed query shape — filter, search, sort, limit, project, in that order. Making that shape planable required a scan leaf that is a real TableProvider, search nodes that are real logical nodes, and a lowering stage that reads only the plan; those three together also make the fixed shape unnecessary. A DataFrame can put a search anywhere and then join, aggregate, or window the result with anything DataFusion can express.

Also adds a planning benchmark, which measures the pipeline with no I/O left in it. It compares both read paths in one process, so Scanner gains a #[doc(hidden)] create_plan_logical for the bench to call — the last PR deletes it along with the path it was there to compare against.

---

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}