Fixes block-scope run-length decode one-past-the-end memory access into smem TempStorage

[elstehle]

Description

Closes #591.

Each thread decodes up to DECODED_ITEMS_PER_THREAD items at a time.
If a thread is assigned to the last run, we make sure that it won't proceed with the next run by setting its one-past-the-end of its current run to correspond to DECODED_ITEMS_PER_THREAD.

    // If this thread is getting assigned the last run, we make sure it will not fetch any other run after this
    DecodedOffsetT assigned_run_end = (assigned_run == BLOCK_RUNS - 1)
                                        ? thread_decoded_offset + DECODED_ITEMS_PER_THREAD
                                        : temp_storage.runs.run_offsets[assigned_run + 1];

However, in the last iteration of their loop, i.e., after having decoded their last of the DECODED_ITEMS_PER_THREAD items, threads enter again the conditional to fetch the next run, if applicable, even if they exit the loop right after, which would result in one-past-the-last item access into temp_storage.runs.run_offsets.

    for (DecodedOffsetT i = 0; i < DECODED_ITEMS_PER_THREAD; i++)
    {
      decoded_items[i] = val;
      item_offsets[i]  = thread_decoded_offset - assigned_run_begin;
      if (thread_decoded_offset == assigned_run_end - 1)
      {
        assigned_run++;
        assigned_run_begin = temp_storage.runs.run_offsets[assigned_run];
        ...
        val                = temp_storage.runs.run_values[assigned_run];
      }
      thread_decoded_offset++;

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.