perf(compute): use typed memo insertion for numeric hashes

[fallintoplace]

Summary

• Use typed memo insertion in the shared numeric hash path.
• Keep the existing normalized uint8, uint16, uint32, and uint64 mappings.
• Avoid the per-value interface call to MemoTable.GetOrInsert.
• Add coverage for int32, int64, float32, float64, and both null modes.
• Add a numeric dictionary_encode benchmark.

Benchmark

Apple M1 Pro. Compared with main at 6b039a76. The benchmark uses 65,535 values, 100 unique values, 500ms per sample, and 5 samples. The table shows median results.

type	main	branch	change	allocs/op
int32	1.168 ms	0.989 ms	15.3% faster	56 -> 56
int64	1.138 ms	1.016 ms	10.8% faster	56 -> 56
float32	1.662 ms	1.114 ms	33.0% faster	64,935 -> 56
float64	1.707 ms	1.251 ms	26.7% faster	64,935 -> 56

Command:

go test ./arrow/compute -run '^$' -bench '^BenchmarkDictionaryEncodeNumeric$' -benchmem -benchtime=500ms -count=5

Checks

• go test ./arrow/compute/...
• go test -race ./arrow/compute/internal/kernels ./arrow/compute -run '^(TestDictionaryEncodeNumericTypes|TestDictionaryEncode|TestDictionaryEncodeArrayWithSmallExecChunkSize|TestDictionaryEncodeResizesMemoTable|TestDictionaryEncodeStateResetAfterFinalize)$' -count=1
• go vet ./arrow/compute/internal/kernels
• git diff --check