ARROW-5588: [C++] Better support for building union arrays

[bkietz]

• Simplify DenseUnionBuilder
• Add SparesUnionBuilder
• MakeBuilder can now produce a {Sparse,Dense}UnionBuilder
• ArrayFromJSON can now produce union arrays

[wesm]

[ RUN      ] TestDictionary.ListOfDictionary
WARNING: Logging before InitGoogleLogging() is written to STDERR
F0702 23:55:33.778414  6046 array.cc:287]  Check failed: list_type_->value_type()->Equals(data->child_data[0]->type)
*** Check failure stack trace: ***
/buildbot/AMD64_Conda_C__/cpp/build-support/run-test.sh: line 97:  6046 Aborted                 (core dumped) $TEST_EXECUTABLE "$@" 2>&1
      6047 Done                    | $ROOT/build-support/asan_symbolize.py
      6048 Done                    | ${CXXFILT:-c++filt}
      6049 Done                    | $ROOT/build-support/stacktrace_addr2line.pl $TEST_EXECUTABLE
      6050 Done                    | $pipe_cmd 2>&1
      6051 Done                    | tee $LOGFILE
/buildbot/AMD64_Conda_C__/cpp/build/src/arrow

[bkietz]

I'll disable that test until I can address the issue in ListBuilder

[fsaintjacques]

Would it simplify to require for children.size() the same as max_type_code()? If you're not, you're asking for trouble. The user can always fill with NullBuilder.

[bkietz]

I can let children_[i] be the builder for type_id == i instead of the builder which will finish into child_data[i]. The latter seems like it might be an implicit contract though; I'll investigate

[fsaintjacques]

This might be unnecessary if we maintain an invariant on children_.

[bkietz]

@fsaintjacques Travis is green https://travis-ci.org/bkietz/arrow/builds/556909415

[fsaintjacques]

LGTM, some small changes and nothing serious. Good job on simplifying the indirection in the builders.

[fsaintjacques]

ToString? Applies to other reference.

[bkietz]

This was a hack to get more info out of the assertion failure than "types are unequal".

[fsaintjacques]

Ok, it passes locally with lint, so I'll proceed to merge.