Add AVX2 kernels for Q4_K and Q6_K

[giankev]

What does this PR do?

Adds AVX2 + FMA implementations for the Q4_K and Q6_K fused dot-product kernels on x86.

The existing NEON and scalar implementations are preserved as fallbacks. AVX2 is enabled at compile time when both __AVX2__ and __FMA__ are available.

On an Intel Core i7-10510U with TinyLlama 1.1B Q4_K_M, generation improved from approximately 1.4 tok/s to 4.2–4.3 tok/s, around a 3x end-to-end speedup.

Type of change

• Bug fix
• New feature
• Performance improvement
• Documentation update
• Refactoring (no behavior change)

Testing

• Tested on x86-64 (Linux/macOS/Windows)
• Tested on ARM64 (Raspberry Pi)
• Tested with TinyLlama 1.1B Q4_K_M

Test command:

./picolm ./models/tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf \
  -p "The capital of Italy is" \
  -n 100 \
  -t 0 \
  -j 4

Output:

Rome.

2. Cities in Italy: Florence, Rome, Venice, Milan, Naples, Turin, Bologna, Verona, Palermo, and Genoa.

...

Prefill: 6 tokens in 1.25–1.59s
Generation: 101 tokens in 23.36–24.02s (4.2–4.3 tok/s)
Total: 24.64–25.61s
Memory: 45.17 MB runtime state (FP16 KV cache)

Additional validation:

• Greedy output matches the scalar implementation.
• Q6_K passed 1,000 randomized scalar-vs-AVX2 dot-product comparisons.
• Five benchmark runs were performed for both baseline and AVX2 builds.
• Baseline generation: 70.97–71.47s (~1.4 tok/s).
• AVX2 Q4_K + Q6_K generation: 23.36–24.02s (~4.2–4.3 tok/s).
• perf profiling showed the expected bottleneck migration from Q4_K to Q6_K after optimizing Q4_K.

Checklist

• Code compiles without warnings (make native)
• No new dependencies added
• Memory usage not increased (check stderr output)
• Works with --json mode (if touching generation/sampling)