[FEA] Support batch construction of strings columns

[ttnghia]

In our framework, we often see this pattern:

for(auto& output_pairs: output_pair_buffers) {
    output.emplace_back(cudf::make_strings_column(output_pairs, stream, mr));
}

That means there are a lot of calls to cudf::make_strings_column, involving a lot of stream synchronization (at least two stream syncs per call: one when generating offsets and one when generating bitmask). In case we have 10, 20 30 etc output columns, having such high count number of stream syncs is very inefficient.

We can do better by implementing a batch construction for strings columns, deferring stream syncs until absolutely necessary. There may be just one stream sync needed for the entire process.

The next optimization level could be to fuse the involved kernels together: instead of calling a separate kernel for constructing each column, we call just one kernel for all columns. For example, calling one valid_if kernel for generating nullmask for all columns, one copy kernel for generating chars data of all columns etc.

[ttnghia]

These are somewhat a similar observation and could need a similar solution:

• [FEA] Improve cudf::gather scalability as number of columns increases #13509
• [TASK][JNI] Investigate train of null_count after explode #11923