perf(parquet): encode RLE levels in batches

[fallintoplace]

What changed

• add a batch path for repetition and definition level encoding
• scan repeated runs once instead of calling the scalar encoder for every level
• keep the scalar path around literal boundaries
• preserve the existing encoded bytes across split streaming calls
• add level encoder and scalar-versus-batch benchmarks

Benchmarks

Apple M1 Pro:

RLE workload	Before	After	Change
65,536 all defined	151,051 ns/op	31,143 ns/op	-79%
65,536 mostly defined	324,228 ns/op	252,108 ns/op	-22%
65,536 alternating	339,128 ns/op	341,149 ns/op	+0.6%

The alternating case is the fallback path. Allocations remain at zero.