High-throughput search and position hot-path optimization

[KnightStackSys]

Performance-only branch from current main focused on NPS and hot-path cost.

Search hot paths

• throttle wall-clock polling to every ~1024 nodes instead of repeatedly per move
• defer partner-in-check detection and HCE evaluation until NMP is otherwise eligible
• generate tactical qsearch moves directly instead of generating all moves then filtering/copying
• remove unused/repeated qsearch check calculations
• merge quiet/capture failure scratch storage into one buffer
• avoid value-initializing large per-node scratch arrays
• at non-PV nodes, avoid sorting the entire move list; score cheaply and lazily select only the strongest first 16 moves as the search actually asks for them
• defer non-PV givesCheck() work until a move is actually considered instead of evaluating check status for the whole move list up front
• preserve the existing full-ordering path at PV nodes

Faster capture ordering

• remove full four-player SEE from the move-ordering score for every generated capture
• use cheap MVV/LVA plus capture history for capture and promotion ordering
• keep SEE where it affects actual search decisions (seeGE pruning)
• avoid pointless givesCheck() calls for captures/promotions when their ordering score does not consume the check bonus

Move generation / storage

• make plain Move scratch buffers trivially default-initialized while preserving Move{} as the null move
• avoid clearing the 512-entry MoveList backing array at construction
• add direct capture/promotion/en-passant tactical generation for qsearch

14-bit compact sliding attacks

• maintain compact occupancy for every rank, file and both diagonal families inside Position
• use one universal 14 x 2^14 attack-mask table
• expand 14-bit attack masks back to the 3x64 board through two 7-bit lookup tables
• encode missing cross-board corner cells as permanent blockers, preventing diagonal attacks from jumping across board holes
• use compact lookup attacks for normal move generation, tactical move generation, HCE mobility and current-position king/check detection
• keep the original ray attack implementation for hypothetical occupancies used by SEE, givesCheck() and castling safety
• randomized property tests cross-check lookup attacks against the original ray implementation before/after make, undo and null moves
• benchmark reports ray vs compact-line queen attack throughput directly

Position / board hot paths

• combine both enemy colors into a single slider attack calculation for arbitrary-occupancy isAttackedByEnemy / attackersToEnemy
• fast-path castling-right updates for the common case where no king/rook is involved
• skip unchanged castling/alive-state Zobrist work
• use no-Zobrist/no-king-state board primitives during undo, then restore reversible state from StateInfo

SEE / TT hot paths

• avoid zeroing the full SEE gain array on every invocation
• use the captured piece already encoded in generated moves instead of rediscovering it from the board
• avoid unnecessary capture-square lookup for normal SEE captures
• stop clearing/default-initializing TTProbe output on every probe/miss; callers use the returned hit boolean

No pruning thresholds, LMR/NMP parameters, evaluation weights, or move legality rules are intentionally changed. Timer polling is intentionally coarser by up to roughly 1024 searched nodes.

The newest move-ordering changes are intentionally more aggressive and may alter node count/Elo, so benchmark both raw NPS and time-to-depth, then run an SPRT before merge.

[KnightStackSys]

This PR significantly improves IronPhoenix's search throughput through compact 14-bit sliding attack tables, faster move generation, reduced search overhead, cheaper capture ordering, lazy non-PV move picking, optimized SEE/TT hot paths, and faster make/undo operations.