[FEA] Fix improper usage of relaxed constexpr in cudf spans

[PointKernel]

Is your feature request related to a problem? Please describe.
Related to #7795

While working on #17746, we noticed that some libcudf code depends on relaxed constexpr in both span and 2dspan (#17746 (comment)), e.g.

cudf/cpp/src/io/parquet/writer_impl.cu

Line 1971 in bbf4f78

ck.fragments = &row_group_fragments.device_view()[c][f];

The above code is trying to access a device span from the host side and this should be done alternatively.

Describe the solution you'd like
For both span and 2dspan, the [] operator should be restricted to either host-only or device-only usage. However, it is valid to obtain a pointer to device storage on the host and vice versa. Therefore, instead of using the [] operator, users should access the storage pointer through data(). In the case of 2dspan, flat_view serves a similar purpose as data does in 1D spans. To align with this naming pattern, we should introduce a new API called flat_index.

CUDF_HOST_DEVICE constexpr size_type flat_index (size_type, row, size_type column) const noexcept {
  return row * this->size().second + column;
}

This approach allows us to address the aforementioned improper usage and make similar adjustments in other instances where the [] operator is misused.

Describe alternatives you've considered

Alternatively, we could follow the STL approach by introducing a mapping to our 2dspan for calculating the "flat index." However, since our 2dspan is not designed as a general-purpose solution like mdspan, it might be more appropriate to refactor the existing flat_view to align with data_handle if we choose this direction. That said, replicating the full complexity of a general-purpose implementation may not worth the effort.