perf(cpu): eliminate hot path index math + add SIMD threshold (#81, #90)

[kolkov]

Summary

Closes #81 (all 8/8 items now complete) and #90 (SIMD min slice threshold).

CPU hot paths (#81) — 3 remaining items fixed

• Reduce: replaced per-element N-D coordinate decomposition with innerOuter block iteration (same pattern as Gather/Chunk fixes)
• Scatter-add: replaced per-element make([]int, ndim) allocation with pre-allocated buffer — allocs O(N) → O(1)
• Batch broadcast matmul: precompute batch offset table — O(1) lookup vs O(ndim) per batch element

SIMD threshold (#90)

• Added simdMinLen = 32 guard to all 47 SIMD dispatch points
• Below 32 elements, scalar loop runs — avoids function-pointer overhead exceeding SIMD savings
• Threshold covers 2 full AVX-256 vectors (2x8) and 2 AVX-512 half-widths (2x16)
• Benchmark + unit test suite added

Test plan

• go build ./...
• go test -short ./internal/backend/cpu/... — 337 tests pass
• golangci-lint run — 0 issues
• Benchmark suite: 9 hot path + 24 threshold benchmarks added

Closes #81
Closes #90

[codecov]

Codecov Report

❌ Patch coverage is 32.55814% with 58 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/backend/cpu/ops_int.go	0.00%	0 Missing and 12 partials ⚠️
internal/backend/cpu/scatter_add.go	0.00%	10 Missing ⚠️
internal/backend/cpu/ops_float64.go	0.00%	0 Missing and 8 partials ⚠️
internal/backend/cpu/ops_float32.go	12.50%	0 Missing and 7 partials ⚠️
internal/backend/cpu/math.go	0.00%	5 Missing ⚠️
internal/backend/cpu/batchmatmul.go	73.33%	4 Missing ⚠️
internal/backend/cpu/reduce.go	80.00%	4 Missing ⚠️
internal/backend/cpu/exp.go	0.00%	0 Missing and 2 partials ⚠️
internal/backend/cpu/relu.go	0.00%	2 Missing ⚠️
internal/backend/cpu/sigmoid.go	0.00%	1 Missing and 1 partial ⚠️
... and 1 more

📢 Thoughts on this report? Let us know!