sqlite3: parse the file header and cell pointer array with combinators

[Hugoberry]

Three structural changes, no behaviour change:

• The 100-byte file header is one hoisted BinaryFormat.Record over a 100-byte slice instead of four magic-offset reads (UInt2(16), Byte(18), Byte(20), UInt4(56)). A length guard keeps the "Not a SQLite 3 database file" error for files too short to hold a header, which Binary.Range would otherwise fail on differently.
• The cell pointer array is contiguous, so it is read with one slice and one BinaryFormat.List rather than ncells separate slice+parse calls.
• DecodeValue dispatches serial types 0-9 through a hoisted list of readers instead of a ten-branch if-chain, so text and blob no longer fall through every integer comparison first. Reserved types 10/11 now error directly rather than reaching the same error via SerialSize.

DecodeRecord was left alone on purpose. Its acc & {x} appends are O(ncols^2) per row, but converting that walk to a List.Generate cursor regressed measurably in 2026-07 and again now: List.Generate costs two closure calls and two state-record allocations per column on every row, and returns a lazy list that must then be buffered before the INTEGER PRIMARY KEY fixup can index it. A comment there records why, so the next reader of that code does not have to rediscover it.

Measured neutral on tests/perf sqlite3-int-decode (5 passes, rotated order, medians within noise). Full PQTest suite green.