bench: cover small run-index selections and full scans

[sunchao]

Why are the changes needed?

Which issue does this PR close?

Related to #10846; this benchmark-only PR does not close the performance issue.

Rationale for this change

RunEndBuffer::get_physical_indices can keep visiting physical runs after it has mapped all requested logical indices. Small selections from large backing buffers need dedicated coverage, alongside full-scan controls, before changing the iteration strategy.

What changes were proposed in this PR?

What changes are included in this PR?

Add a Criterion benchmark for two-element prefix selections, two-element selections from a sliced buffer, and all-index selections. Each case runs with 1,024 and 1,048,576 physical runs. Input construction stays outside the timed loop. This separate benchmark PR follows the contributor guide's performance workflow.

Are there any user-facing changes?

None. This PR adds benchmarks only.

How was this PR tested?

Are these changes tested?

• Ran all six benchmark cases on the unchanged production implementation at 900ec3ee38276ab651e210c4a85b38f8a8a61bcf, with this benchmark added.
• Formatting and affected-crate all-target/all-feature Clippy passed.
• Local measurements are a development check, not an automated-runner or end-to-end query-performance claim.

AI assistance: Codex generated the benchmark and PR text and ran the stated local checks. This does not claim a separate human review.