perf(arrow/array): batch run-end encoded appends

[fallintoplace]

What

• Batch RunEndEncodedBuilder.AppendNulls into one physical null value and one run.
• Batch RunEndEncodedBuilder.AppendEmptyValues into one physical empty value and one run.
• Keep zero and negative counts as no-ops.
• Preserve the empty-value state used by JSON decoding.

The old implementation called the scalar append method once per logical value. This change keeps the same API while avoiding the repeated physical values and run ends.

Benchmark

Apple M1 Pro, 65,536 rows, int32 run ends and int32 encoded values:

Case	Before	After	Speedup
Nulls	~1.0-1.6 ms, 1.16 MB, 55 allocs/op	~1.0-1.5 us, 1.7 KB, 19 allocs/op	~1,250x
Empty values	~1.2-2.7 ms, 1.16 MB, 55 allocs/op	~1.2-1.5 us, 1.7 KB, 19 allocs/op	~1,750x

The benchmark covers int16, int32, and int64 run-end types, int32 and string encoded values, and multiple batch sizes.

Tests

• go test ./arrow/array -count=1
• go test ./... -run '^$' -count=1 -p 2

[zeroshade]

Reviewed at fa9d776. The bulk append implementation preserves logical behavior while reducing repeated null/empty appends to a single physical run. The full arrow/array suite, race tests, semantic equivalence probes, boundary coverage, formatting, vet, and diff hygiene all passed locally.