Add PhoenixNet NNUE v1

[KnightStackSys]

Adds the first working IronPhoenix NNUE baseline for 4-player Teams chess.

Engine

• PhoenixNet v1 NNUE feature mapping/inference
• 17 king buckets
• own-king + partner-king streams
• side-to-move-team-relative output
• HCE fallback when no net is loaded

MultiPV

• dedicated root MultiPV search
• original PVS/aspiration path unchanged when MultiPV=1
• UCI setoption name MultiPV value N, range 1..32
• standard multipv N info output

Dataset generation

• multi-worker HCE/search self-play
• exact C++ NNUE::featureIndex() sparse features
• controlled MultiPV exploration
• randomized opening plies
• feature-input deduplication
• append-safe native IPDATA1 files

Graceful stop + resume

• Ctrl+C requests a graceful stop instead of killing the dataset writer
• active teacher searches are stopped, workers exit, and the current record count is committed before return
• progress is reported every 10 saved positions
• --resume --positions N treats N as the desired TOTAL dataset size; e.g. 50k existing + --resume --positions 500000 generates the remaining 450k
• --append --positions N keeps its existing meaning: add N new positions
• resume/append scan and preload existing dedup fingerprints
• interrupted older datasets with stale header counts are scanned for complete trailing records; complete records are recovered and only an incomplete final record is discarded
• recovered record counts are written back to the dataset header before generation continues

Training tooling

• nnue/model.py
• nnue/train.py reads .ipd directly
• game-level train/validation splitting
• vectorized EmbeddingBag packing
• nnue/export.py

Current optimization stage

PhoenixNet v1 intentionally performs full feature-transformer refreshes as a correctness baseline. Incremental accumulators, quantization, and SIMD are later optimization stages.

[KnightStackSys]

This PR will also add MultiPV support