perf(parquet): use typed memo insertion for fixed-width dictionaries

[fallintoplace]

What changed

• Use BinaryMemoTable.InsertOrGet for FLBA dictionary Put and PutDictionary.
• Use the same typed insertion path for DictInt96Encoder.Put.
• Keep empty Put calls as no-ops.
• Add Int96 dictionary round-trip coverage and FLBA/Int96 benchmarks.

Why

These encoders already use the binary memo table, but still called GetOrInsert through the generic interface for every value. The typed entry point avoids the extra interface conversion. Dictionary order and encoded output stay unchanged.

Benchmark

Apple M1 Pro, macOS arm64, Go 1.26.3. 65,535 values and 100 unique values, median of five runs.

Case	Before	After	Change
FLBA, width 16	4.59 ms	3.09 ms	33% faster
Int96	5.41 ms	3.87 ms	29% faster
FLBA allocations	131,117	65,581	50% fewer
Int96 allocations	196,655	131,118	33% fewer

Allocated bytes also fell by about 42% for FLBA and 33% for Int96.

Tests

• go test ./... with Arrow and Parquet test data
• go test ./parquet/internal/encoding/...
• go test -race ./parquet/internal/encoding/...
• go vet ./parquet/internal/encoding/...
• go test -tags noasm ./parquet/internal/encoding/...

[zeroshade]

The typed memo insertion preserves fixed-width value ownership, sliced offsets, and Int96 dictionary encoding. Focused dictionary encoding tests passed (go test -p 2 -count=1 ./parquet/internal/encoding -run Dict).