feat(ai): add stable softmax, normalization and activation kernels

[damian123]

Summary

This is a focused three-file delta on top of #34:

• Vectorisation/VecX/neural_kernels.h
• VectorTest/TestNeuralKernels.cpp
• VectorTest/CMakeLists.txt

Review the incremental #34 to #35 diff.

It adds AVX2 VecF8F preprocessing kernels for:

• softmax, log-softmax, and log-sum-exp
• LayerNorm and RMSNorm
• ReLU, Leaky ReLU, Sigmoid, Tanh, approximate GELU, and SiLU
• clipping and bias-plus-activation
• exact in-place and out-of-place operation

Numerical contracts

• Softmax-family kernels subtract the logical maximum before exponentiation. Their sums visit only the first n values, in index order, and accumulate in double precision; SIMD padding never participates.
• LayerNorm uses population variance and preserves the residual of its double-precision mean when centering float32 values. RMSNorm uses mean square. Both apply finite, positive epsilon inside the square root.
• Inputs and optional affine arrays must be finite. NaN and infinity are rejected with std::invalid_argument rather than producing accidental non-finite results.
• Leaky ReLU alpha must be finite and non-negative. Clip bounds must be finite and ordered.
• Buffers are contiguous float32 arrays. Caller alignment is not required, and logical sizes below, at, and above the eight-float SIMD width are supported.
• Exact input/output aliasing is supported. Arbitrary partial overlap is outside the contract.
• Empty log-sum-exp returns negative infinity; other empty-buffer operations are no-ops after parameter validation.
• Reduction order is deterministic. Cross-compiler floating-point results are compared to independent scalar-double references with stated tolerances, not bit identity.

Verification

• 15 deterministic focused tests cover width-8 boundaries (1, 7, 8, 9, 15, 16, 17), deliberately unaligned buffers, output canaries, exact aliases, extreme finite logits, non-finite rejection, zero variance, epsilon-dominant inputs, and scalar-double error metrics.
• GCC 13.3 Release: full build including examples; 8/8 CTest entries passed.
• Clang 18.1 Release: 4/4 focused CTest entries passed.
• GCC 13.3 Debug with ASan/UBSan: 4/4 focused CTest entries passed.

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DDR3_BUILD_TESTS=ON -DDR3_BUILD_EXAMPLES=OFF -DDR3_ISA=AVX2
cmake --build build --config Release --target VectorTest
ctest --test-dir build -C Release -R 'Softmax|LayerNorm|RMSNorm|Activation' --output-on-failure

Non-goals

• No training loop.
• No autograd or gradient tape.
• No tensor abstraction, model loader, or general inference runtime.
• No AVX-512-specific path until the maintainer opts into that direction.
• No universal performance claim.

This PR remains a draft pending #34 and maintainer guidance on final component placement.