LSX/LASX NNUE paths optimizes
Optimizes the LoongArch NNUE SIMD implementation, avoiding permutation and shuffle operations in `ClippedReLU` and `SqrClippedReLU`.
Adds LSX path that widens and processes both halves of each signed-byte weight vector.
Passed STC:
LLR: 2.94 (-2.94,2.94) <0.00,2.00>
Total: 50528 W: 13459 L: 13130 D: 23939
Ptnml(0-2): 182, 5438, 13706, 5745, 193
https://tests.stockfishchess.org/tests/view/6a607ab325028d004c922829
closes https://github.com/official-stockfish/Stockfish/pull/7005
No functional change