[core] Reduce OSS requests for dense manifest sidecar selections

[JingsongLi]

Purpose

Follow up on #988: a scan that uses a manifest sidecar can issue several range GETs for one manifest even when most of its blocks are selected. For an approximately 8 MiB manifest, this can turn one full GET into a sidecar GET plus two or more range GETs.

Changes

• Estimate selected manifest bytes and the number of range reads after the existing 4 MiB coalescing rule.
• Read the full manifest when doing so saves range requests and the transfer increase is bounded by the range count, capped at 4x. Keep sparse selections on the range-read path and empty selections header-only.
• Use the same range planner for the estimate and the actual selected read.

For an 8 MiB manifest with two planned ranges and at least 4 MiB selected, the manifest body now uses one full read rather than two range reads.

Tests

• cargo test -p paimon spec::manifest_sidecar::tests --lib (8 passed)
• cargo test -p paimon table::table_scan::tests --lib (109 passed)
• cargo clippy -p paimon --lib --tests -- -D warnings
• cargo fmt --all -- --check

No storage format change.

[leaves12138]

LGTM. Reviewed e57d695 and found no blocking issues. Checked that the estimate and selected read share the same range planner, the full-read policy respects the request-count threshold and 4x amplification cap, empty selections remain header-only, and the scan callers still apply entry-level filtering after a full-manifest fallback.

Local validation: all 8 existing manifest-sidecar tests and 109 table-scan tests passed. Three additional review-only tests also passed: exact threshold/4x-cap boundaries; request counting through the OSS backend against a local HTTP endpoint (one manifest-body GET for dense or large-contiguous selections, two range GETs for the sparse case, zero body requests for an empty selection); and scan-result equivalence with sidecars disabled across dense, sparse and empty row-range selections. All 14 GitHub checks were successful for this head. The extra tests were kept in the isolated review checkout; no changes were pushed.