TBB implementation of reduce_by_key incompatible with transform_output_iterator

[harrism]

As part of NVIDIA/thrust#1805 I added a new test that combines reduce_by_key with transform_output_iterator. This fails to compile due to this line:

https://github.com/NVIDIA/thrust/blob/d641d63802d08c17bde9498677be300cb4277bdf/thrust/system/tbb/detail/reduce_by_key.inl#L329

I have disabled that test for the TBB backend using the Cmake macro thrust_declare_test_restrictions(). Once this issue is fixed that test should be enabled for TBB.

[gevtushenko]

I can reproduce this issue. The algorithm looks as follows:

1. Input sequence is divided between P threads
2. Each thread reduces it's chunk of data in reverse until a key changes
3. Accumulated value is stored in carries array
4. The rest of the chunk is processed by sequential reduce by key
5. When all threads complete, we combine result of the incomplete chunks (carries) with a successor tile's partial result

The issue is related to the final step. Successor tile's partial result is written into values_result, which is fine. But we are trying to read value_results to combine it with carries[i], which is an issue since values_result is an output iterator.

We'd probably have to store partial result for the first key in the chunk in a temporary storage as well.