perf(compute): vectorize numeric to boolean casts

[fallintoplace]

What does this PR do?

• Adds an ARM64 NEON fast path for numeric to boolean casts.
• Reuses the existing NEON array != scalar comparison kernel to produce the output bitmap.
• Covers int32, uint32, int64, uint64, float32, and float64.
• Keeps the scalar path for narrow integers, noasm builds, unsupported CPUs, and appengine.
• Adds coverage for NaN, signed zero, infinities, sliced input, and bitmap tails.
• Adds a benchmark matrix for the supported types, sizes, and zero distributions.

Benchmark

Apple M1 Pro, GOMAXPROCS=1, compared with -tags noasm.

• 32-bit types: about 9x faster at 65K values and about 14x faster at 1M values.
• 64-bit types: about 6x faster at 65K values and about 8x faster at 1M values.
• Small arrays keep the same allocation count and retain the scalar tail path.

Tests

• go test ./arrow/compute ./arrow/compute/internal/kernels
• go test -tags noasm ./arrow/compute ./arrow/compute/internal/kernels
• amd64 compile check for the compute package noasm fallback

[zeroshade]

Verified locally on Apple M1: the NEON path is exercised natively and passes, including NaN, signed zero, infinities, sliced inputs, and bitmap tails; noasm parity passes. Reusing the existing comparison kernel keeps the semantics aligned with the scalar isNonZero path.

---

This review was drafted by an AI-assisted tool and
confirmed by an Apache Arrow Go maintainer.