ARROW-6042: [C++][Parquet] Add Dictionary32Builder that always returns 32-bit dictionary indices

[wesm]

Without this, DictionaryArrays produced by different parts of a Parquet file could have different index types depending on the cardinality of each decoded part.

Refactors DictionaryBuilder (which uses AdaptiveIntBuilder) and Dictionary32Builder (which uses Int32Builder) to have a common base class.

[codecov-io]

Codecov Report

Merging #4956 into master will increase coverage by 1.62%.
The diff coverage is 97.4%.

@@            Coverage Diff             @@
##           master    #4956      +/-   ##
==========================================
+ Coverage   87.49%   89.11%   +1.62%     
==========================================
  Files         998      721     -277     
  Lines      141784   101687   -40097     
  Branches     1418        0    -1418     
==========================================
- Hits       124058    90623   -33435     
+ Misses      17364    11064    -6300     
+ Partials      362        0     -362

Impacted Files	Coverage Δ
cpp/src/parquet/encoding-test.cc	100% <ø> (ø)	⬆️
cpp/src/parquet/encoding.h	97.82% <ø> (ø)	⬆️
cpp/src/parquet/column_reader.cc	88.93% <ø> (ø)	⬆️
cpp/src/arrow/array-dict-test.cc	95.52% <100%> (+0.13%)	⬆️
cpp/src/parquet/arrow/arrow-reader-writer-test.cc	93.71% <100%> (ø)	⬆️
cpp/src/parquet/encoding.cc	93.73% <100%> (ø)	⬆️
cpp/src/arrow/array/builder_dict.h	88.57% <95.12%> (+0.33%)	⬆️
cpp/src/arrow/array/builder_adaptive.h	93.75% <0%> (-3.13%)	⬇️
go/arrow/ipc/writer.go
... and 278 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 38b0176...5210b30. Read the comment docs.

[xhochy]

+1, LGTM

[pitrou]

+1. I'm curious, does this make parquet conversion faster?

[wesm]

@pitrou I haven't done comprehensive benchmarks yet but the main benefit is actually memory use (see example of user running into runaway memory usage on a highly encoded file at https://issues.apache.org/jira/browse/ARROW-5993). Performance should be better, too. I'll run benchmarks to illustrate

[wesm]

BTW @pitrou this patch is also necessary if you want to add dictionary encoding to the CSV reader (because each chunk may have a different dictionary cardinality)