perf(compute): add ARM64 NEON comparison kernels

[fallintoplace]

Summary

• Added ARM64 NEON kernels for primitive comparisons.
• Covered int32, uint32, int64, uint64, float32, and float64.
• Kept narrow integer types and unsupported builds on the generic path without the NEON wrapper.
• Removed redundant 64-bit packing shifts from the hot loops.
• Handled array/array, array/scalar, scalar/array, bitmap offsets, tails, NaNs, and unsigned boundary values.

Benchmark

Apple M1 Pro, 65,536 int64 values, nullprob=0.000000, 3 runs with -benchtime=1s. The generic baseline uses -tags noasm. Values below are medians.

Case	NEON	Generic	Speedup
Array/scalar	29.6 us/op	76.3 us/op	~2.6x
Array/array	29.3 us/op	313.6 us/op	~10.7x

Native uses 29 allocs/op versus 30 for the generic baseline in both cases.

Tests

• go test ./arrow/compute/internal/kernels ./arrow/compute -count=1
• go test -tags noasm ./arrow/compute/internal/kernels ./arrow/compute -count=1
• go test -race ./arrow/compute/internal/kernels ./arrow/compute -count=1
• go vet ./arrow/compute/internal/kernels
• Linux ARM64 and AMD64 cross-builds for the touched packages
• Differential ARM64 coverage for offsets 0-7 and lengths through 1024