perf(simd): add minimum slice-length threshold before dispatching to SIMD

[bennibbelink]

Context

PR #72 dispatches to SIMD via function pointers whenever simdXxx != nil. There is currently no lower bound on slice length before invoking the SIMD path.

Problem

For small slices, the overhead of a function-pointer call plus SIMD register setup may exceed the cost of the scalar loop. See #72 (review) for more context

Suggested Fix

1. Benchmark explicitly with small slice sizes (1, 3, 7, 15 elements) to find the empirical crossover point.
2. Add a threshold guard in each dispatch function, e.g.:

const simdMinLen = 32  // tune after benchmarking

if simdAddInplaceFloat32 != nil && len(a) >= simdMinLen {
    simdAddInplaceFloat32(a, b)
    return
}
// fall through to scalar

The exact value of simdMinLen should be determined per-operation and
per-dtype

[kolkov]

Implemented in v0.9.23 (PR #150): simdMinLen=32 guard on all 47 SIMD dispatch sites. Benchmark + unit test suite included.