Add aarch64 SIMD dispatch paths across all CPU kernel modules to enable
ARM NEON acceleration on Apple Silicon and other ARM64 platforms.

Previously, SIMD optimizations were limited to x86-64 (AVX2/AVX-512).
This change extends the dispatch pattern to include aarch64, allowing
f32/f64 operations to use NEON instructions on ARM processors.

Modified kernel modules:
- Binary operations (add, sub, mul, div)
- Comparison operations (eq, lt, gt, etc)
- Unary operations (neg, abs, sqrt, exp, relu)
- Activation functions (sigmoid, silu, gelu, leaky_relu, elu)
- Reduction operations (sum, mean, max)
- Normalization (layer_norm, rms_norm)
- Indexing (masked_select, masked_fill, masked_count)
- Conditional select (where)

All changes follow the existing dispatch pattern: try x86-64 SIMD,
try aarch64 SIMD, fall back to scalar implementation.