fix: eliminate per-element heap allocs in is_in kernel for binary types

[zeroshade]

Summary

• Add BinaryMemoTable.ExistsDirect that inlines the hash table probe loop, avoiding the closure in HashTable.Lookup that causes val []byte to escape to the heap
• Add isInBinaryDirect specialized kernel path that bypasses the visitBinary → VisitBitBlocksShort closure chain by directly iterating with OptionalBitBlockCounter
• Route BinaryDataType dispatch in DispatchIsIn to the new direct path (handles both int32 and int64 offsets)

Motivation

The is_in kernel for binary types allocated once per input element because the []byte value escaped to the heap through a closure chain:

1. visitBinary slices rawBytes[offsets[pos]:offsets[pos+1]] and passes to a callback
2. The callback calls BinaryMemoTable.Exists(v)
3. Exists calls lookup which creates a closure capturing val
4. The closure is passed to HashTable.Lookup, causing escape analysis to move val to the heap

Closes #736

Benchmark (100k rows, 10-element value set)

Metric	Before	After	Improvement
ns/op	4,133,679	923,565	4.5x faster
B/op	2,435,327	33,092	73x less memory
allocs/op	100,075	70	1,430x fewer allocs

All existing TestIsInBinary subtests pass (binary, large_binary, utf8, large_utf8 × all null matching behaviors).

[zeroshade]

CC @laskoviymishka