perf(compute): vectorize mixed 64-bit filters

[fallintoplace]

What changed

• Add AVX2 and ARM64 NEON paths for null-free 64-bit fixed-width filters.
• Compact two groups of four uint64 values from each filter byte with table-driven permutations and exact-size stores.
• Keep the existing path for nulls, unsupported CPUs, short or unaligned inputs, dense masks, and noasm builds.
• Add exhaustive mask tests, guard tests, and int64/float64 benchmarks.

Benchmark

Apple M1 Pro, macOS arm64, Go 1.26.3. Median of five runs. The baseline uses GODEBUG=cpu.all=off.

Case	Baseline	NEON	Speedup
int64, 1M, alternating	3.97 ms	0.45 ms	8.7x
int64, 1M, random50	2.86 ms	1.61 ms	1.8x
float64, 1M, alternating	4.07 ms	0.47 ms	8.6x
float64, 1M, random50	2.92 ms	1.65 ms	1.8x

The mixed-mask cases also go from 33 to 32 allocations per operation. The output allocation is still the main B/op cost.

Tests

• Full go test suite with Parquet test data
• go test ./arrow/compute/...
• go test -tags noasm ./arrow/compute/...
• go test -race ./arrow/compute/internal/kernels ./arrow/compute
• amd64 cross-build of the kernel package