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

[fallintoplace]

What does this change?

• doAppendBinary and doAppendFixedSize still called MemoTable.GetOrInsert for every value.
• The memo is already a TypedMemoTable[[]byte].
• Use InsertOrGet directly, like the numeric hash paths.
• Null handling is unchanged.

Why

This removes the []byte to interface{} call-site conversion from the compute dictionary and unique kernels. It also makes the typed binary memo path reachable from these hot loops.

Benchmark

Command:

go test -p 1 -vet=off ./arrow/compute -run '^$' -bench '^BenchmarkDictionaryEncodeBinary$' -benchmem -benchtime=500ms -count=7

Apple M1 Pro, macOS arm64, Go 1.26.3. The numbers below are medians of 7 runs. Baseline uses GetOrInsert at both call sites. After uses TypedMemoTable[[]byte].InsertOrGet.

Case	Before	After	Change
string, 65,536 values	4,837,326 ns/op	3,151,276 ns/op	34.9% faster
fixed-size binary 16, 65,536 values	4,445,131 ns/op	3,134,294 ns/op	29.5% faster
string B/op	3,444,557	1,871,270	45.7% fewer
fixed-size binary B/op	3,444,128	1,870,749	45.7% fewer
string allocs/op	131,150	65,613	50.0% fewer
fixed-size binary allocs/op	131,146	65,609	50.0% fewer

Coverage

• Existing string and binary dictionary tests cover the variable-width path.
• Added a fixed-size-binary dictionary test.
• Added string and 16-byte fixed-size-binary benchmarks.
• go test -p 1 ./arrow/compute -count=1

No public API changes.