Merge pull request #14 from tphakala/feature/reverse-addsub

Add Reverse and AddSub SIMD operations for f32

Merge pull request #12 from tphakala/feature/butterfly-complex

Add ButterflyComplex for fused FFT butterfly with twiddle multiply

Merge pull request #11 from tphakala/feature/split-complex-ops

Add c64 package and split-format complex operations

Merge pull request #10 from tphakala/feature/f16-half-precision

Add f16 package for half-precision (FP16) SIMD operations

Merge pull request #8 from tphakala/feature/int-to-float-scale

Add Int32ToFloat32Scale for audio PCM conversion

Fix ARM64 NEON sigmoid instruction encodings

Corrected multiple wrong instruction encodings in sigmoidNEON:
- FNEG: 0x6EA0F800 → 0x6EA07C00
- FMIN: 0x4E3BF400 → 0x4EBBF400
- FMAX: 0x4E38F400 → 0x4E3CF400
- FRINTN: 0x4EA19822 → 0x4E218C22
- FCVTZS: 0x4EA1A841 → 0x4EA1B841
- SHL: 0x4F575C21 → 0x4F375421
- ADD: 0x4EA18421 → 0x4EB68421

The original encodings caused SIGILL on ARM64 due to invalid
instruction bytes.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Fix sigmoid to use accurate exp-based computation

Replace fast but inaccurate soft-sign approximation with proper sigmoid
using range reduction and polynomial exp approximation.

Old (wrong): σ(x) ≈ 0.5 + 0.5*x/(1+|x|)  - up to 7.6% error
New (correct): σ(x) = 1/(1+exp(-x))       - float32 precision

Algorithm:
- Clamp input to [-20, 20] to prevent overflow
- Range reduction: exp(-x) = 2^k * exp(r)
- 5-term Taylor polynomial for exp(r)
- Reconstruct via IEEE754 exponent manipulation

Performance: ~17x faster than pure Go (vs ~40x with wrong approximation)
Throughput: 23.8 GB/s on AVX, matching pure Go accuracy.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Merge pull request #5 from tphakala/feature/activation-functions

Add neural network activation functions with SIMD optimizations

Add govet linter to .golangci.yaml and update assembly code for compl…

…ex128 operations

- Added the govet linter to the GolangCI configuration for improved code analysis.
- Updated assembly code in c128_amd64.s to use more descriptive variable names (s_real and s_imag) for clarity.
- Adjusted frame sizes in various functions to ensure proper alignment and memory usage.

Update documentation in doc.go to reflect changes in available operat…

…ions

- Added new arithmetic operations: AddScaled, FMA
- Updated reductions to include DotProductBatch, MinIdx, MaxIdx
- Introduced new statistics functions: StdDev, EuclideanDistance, Normalize
- Enhanced element-wise operations with Reciprocal
- Added AccumulateAdd and CumulativeSum to Audio DSP section
- Improved clarity and organization of the documentation