lib: preserve MxDSSelectAction list extension through Read/Write

[foxtacles]

MxDSSelectAction child lists ('Act_'/'RAND') replace the regular list count with a null-terminated selector string (e.g. "RANDOM_3"), the actual count, and one null-terminated value string per child. Read previously consumed this extension only for the description stream (assuming single-character values) and Write always emitted a bare child count, corrupting select actions in rewritten files.

This retains the raw extension bytes on the object (Object::list_extension_) and re-emits them verbatim on write. CalculateMaximumDiskSize accounts for the extra bytes.

Verified byte-exact roundtrip of the extension on the four retail SI files that use it (COPTER.SI, DUNECAR.SI, JETSKI.SI, RACECAR.SI), and full Read→Write→Read object-tree equality on all 26 retail 1.1 English SI files.

[foxtacles]

Regression verification done against retail 1.1 English data and the isle-portable asset pipeline:

• From-scratch writer unchanged: isle-portable's asset_generator (widescreen/hdmusic/badend/rabbits packs) produces byte-identical output built against master's lib vs this branch.
• Retail rewrite fidelity: Read→Write→Read over all 26 retail SI files compares equal on every object field and every data chunk (including the select-action extension, which now survives verbatim — byte-checked in the four Build SIs that use RAND lists). MxOf count round-trips retail-exact (e.g. ISLE.SI 3708; the engine reads this field to size and fill its offset table — mxdsfile.cpp reads m_lengthInDWords dwords — so preserving it reproduces retail behavior exactly; the previous slot-count value was the deviation).
• Byte-consumption parity: for retail RAND lists (single-character values) the new null-terminated-string parse consumes exactly the same bytes as the old u16 model, so parse alignment is unchanged; for multi-character values or Act_ lists the old model desynced and the new one follows the engine's deserializer layout.
• Boundary safety: a structural audit (headers must not straddle buffer boundaries, non-split MxCh must not cross) reports zero issues over retail files rewritten with size-changed objects — the scenario that corrupted files before the padding fixes. CalculateMaximumDiskSize only grew, so padding can only become more conservative.

One caveat worth knowing: if an object is added to or removed from a file, the preserved MxOf count becomes stale by that delta (previously it was rewritten with the top-level slot count, which was also not the retail semantic). Fresh files with no count read still fall back to the child count.