Write: honor declared sort orders on append and compaction

[philcunliffe]

Summary

Apply a table's declared sort order when writing data files (append) and when compacting, so per-file column bounds are tight and non-overlapping across files. Sorting is currently listed as unsupported.

Current behavior

• README "Supported Features": Sorting ❌.
• src/create.js accepts and records a sortOrder in table metadata (sort-orders / default-sort-order-id), but icebergAppend writes records in input order, and compaction does not enforce the sort order.

Proposed

• Append: order records within each written data file by the table's default-sort-order (or a per-call override).
• Compaction / rewrite: produce sorted output files whose ranges are non-overlapping for the sort key.
• Keep it physical-only: row contents and counts unchanged.

Acceptance criteria

• Data files produced by append and compaction are ordered by the declared sort order.
• After compaction, per-file column bounds for the sort key are non-overlapping.
• Round-trip read returns identical rows (modulo order).

Why it matters

Sorted/clustered layout is what makes file-level and row-group-level pruning (companion issues) effective — without it, file and row-group bounds overlap and pruning can skip very little. Callers can pre-sort a single append's input as a partial workaround, but cross-file and compaction-time sort need engine support.

[philcunliffe]

Part of an Iceberg scan-performance capability set, intended to land together:

• Read pruning: skip data files via partition values + manifest column bounds #20 — file / partition pruning (skip non-matching data files)
• Read pruning: skip parquet row groups via column statistics #21 — row-group pruning (skip non-matching parquet row groups)
• Write: honor declared sort orders on append and compaction #22 — sort order on append / compaction (tight, non-overlapping bounds)

Pruning (#20 / #21) skips non-matching data on read; sort (#22) makes the per-file and per-row-group bounds tight enough for that pruning to skip most of the table. Together they turn a selective WHERE from a full scan into a targeted read.

[philcunliffe]

Completed by #23 (merged as 3edb15b). Append now sorts each written file by the table's default sort order (src/write/sort.js buildSortComparator, applied in prepareAppend) and records the real sort_order_id, with a per-call sortOrderId override threaded through icebergAppend/icebergStageAppend/tx.append. Compaction landed as icebergRewrite (src/write/rewrite.js): reads all live rows with deletes applied, sorts globally, and writes consolidated non-overlapping files via a replace snapshot. Deferred follow-ups noted in the PR: v3 rewrite (must preserve _row_id lineage) and retry-on-conflict for rewrite. (Stayed open because the PR's "Closes #20, #21, #22" only auto-closed the first issue.)