perf(parquet): optimize numeric dictionary insertion

[fallintoplace]

What changed

• use typed memo insertion for int32, int64, float32, and float64 dictionary encoders
• reserve the index buffer for the whole input batch and fill it directly
• keep the existing byte-array and fixed-length byte-array paths unchanged
• add correctness coverage for all four numeric types, including multiple Put calls
• add a numeric dictionary encoding benchmark

Why

The numeric path was calling the generic dictionary insertion method for every value. That added interface and reflection overhead, and the index buffer could grow inside the hot loop.

The new path does the memo lookup with the concrete numeric type and sizes the index buffer once per batch. Dictionary order, duplicate handling, and encoded indices stay the same.

Benchmark

With 65,535 values and 100 unique values on an Apple M1 Pro:

• int32: about 1.21 ms/op to 0.89 ms/op, 26 to 13 allocs/op
• int64: about 1.21 ms/op to 0.92 ms/op, 26 to 13 allocs/op
• float32: about 2.1 ms/op to 1.1 ms/op, 64,905 to 13 allocs/op
• float64: about 2.2 ms/op to 1.2 to 1.4 ms/op, 64,906 to 13 allocs/op

Validation

• go test ./parquet/internal/encoding -count=1
• focused Parquet Arrow dictionary tests
• go vet ./parquet/internal/encoding
• git diff --check