[ntuple] Schema evolution of record fields

[hahnjo]

Description

We currently have no RRecordField::BeforeConnectPageSource that either checks or correctly schema evolves record fields.

Reproducer

#include <ROOT/RNTupleModel.hxx>
#include <ROOT/RNTupleReader.hxx>
#include <ROOT/RNTupleWriter.hxx>

#include <cstdint>
#include <string>
#include <tuple>

void ntuple_tuple() {
  {
    auto model = ROOT::RNTupleModel::Create();
    using TupType = std::tuple<std::int32_t, std::string>;
    auto tup = model->MakeField<TupType>("tup");
    auto writer = ROOT::RNTupleWriter::Recreate(std::move(model), "ntpl", "ntuple_tuple.root");

    *tup = {42, "abc"};
    writer->Fill();
  }

  { // 1
    auto model = ROOT::RNTupleModel::Create();
    using TupType = std::tuple<std::int32_t, std::string, float>;
    auto tup = model->MakeField<TupType>("tup");
    try {
      auto reader = ROOT::RNTupleReader::Open(std::move(model), "ntpl", "ntuple_tuple.root");
    } catch (const ROOT::RException &e) {
      std::cout << e.GetError().GetReport() << "\n";
    }
  }

  { // 2
    auto model = ROOT::RNTupleModel::Create();
    using TupType = std::tuple<std::string>;
    auto tup = model->MakeField<TupType>("tup");
    try {
    auto reader = ROOT::RNTupleReader::Open(std::move(model), "ntpl", "ntuple_tuple.root");
      reader->LoadEntry(0);
      std::cout << "0: string = " << std::get<0>(*tup) << "\n";
    } catch (const ROOT::RException &e) {
      std::cout << e.GetError().GetReport() << "\n";
    }
  }
}

This macro writes a std::tuple<std::int32_t, std::string> and then demonstrates two cases:

1. When trying to read back as a std::tuple<std::int32_t, std::string, float>, the additional float leaf field complains:

No on-disk field information for `tup._2`
At:
  const ROOT::RFieldBase::ColumnRepresentation_t &ROOT::RFieldBase::EnsureCompatibleColumnTypes(const ROOT::RNTupleDescriptor &, std::uint16_t) const

2. With an imposed field type of std::variant<std::string>, the user could expect that the string is read back. Instead it (correctly) throws:

On-disk column types {`SplitInt32`} for field `tup._0` cannot be matched to its in-memory type `std::string` (representation index: 0)
At:
  const ROOT::RFieldBase::ColumnRepresentation_t &ROOT::RFieldBase::EnsureCompatibleColumnTypes(const ROOT::RNTupleDescriptor &, std::uint16_t) const

Additional context

We discussed a conservative approach to check exact identity of the type name.

[hahnjo]

Here it is less clear (to me) what we want to support: For classes, both cases are taken care of - because we have field names that we can add or reorder...