perf: NEON SIMD matmul optimization (~4x speedup)

[dddimcha]

Summary

Replace scalar tensor_gemm() with SIMD-accelerated matmul_neon() in tensor_dense_forward().

Changes

• 1 file changed: kernel/ai/tensor_ops.c (+62, -8 lines)
• Uses existing matmul_neon() from simd_ops.c
• Converts float→Q16.16 fixed-point→SIMD→float
• Graceful fallback to scalar on allocation failure

Benchmark Results (ARM64)

Matrix Size	Scalar	NEON	Speedup
64×64	0.16ms	0.04ms	4.48x
128×128	1.52ms	0.38ms	4.03x
256×256	15.2ms	3.76ms	4.04x
512×512	144ms	36ms	4.00x
1024×1024	1278ms	319ms	4.01x

Testing

• ✅ Compiles for x86_64
• ✅ Compiles for aarch64 (cross-compile)
• ✅ Benchmarked on Apple Silicon

Code Diff

// BEFORE (scalar):
tensor_gemm(in_data, weight_data, out_data, ...);

// AFTER (SIMD):
matmul_neon(in_fixed, weight_fixed, out_fixed, ...);