perf(parquet): decode levels directly into int16 batches

[fallintoplace]

What changed

• decode RLE and bit-packed levels directly into []int16
• count max definition levels during decoding
• remove the intermediate uint64 level batch and its second conversion pass
• add benchmarks for all-defined, mostly-defined, and alternating levels
• check partial value counts when a literal run is truncated

Benchmarks

Apple M1 Pro, -benchtime=300ms -count=5:

RLE workload	Before	After	Change
1,024 all defined	2,137 ns/op	1,495 ns/op	-30%
1,024 mostly defined	5,740 ns/op	5,329 ns/op	-7%
65,536 all defined	61,213 ns/op	24,601 ns/op	-60%
65,536 alternating	120,994 ns/op	89,543 ns/op	-26%

Allocations were unchanged.

[zeroshade]

LGTM. The direct int16 level-decoding path preserves RLE and bit-packed semantics across partial reads, maximum levels, split decoding, and malformed inputs. Package tests, race tests, the full Parquet suite, randomized differential probes, benchmarks, and CI all pass.

---

This review was drafted by an AI-assisted tool and
confirmed by an Apache Arrow Go maintainer. The maintainer
approving this PR has read the findings and signed off. If
something feels off, please reply on the PR and a maintainer
will follow up.

More on how Apache Arrow Go handles maintainer review:
CONTRIBUTING.md.