Fix memset error in nvtext::edit_distance_matrix

[davidwendt]

Description

Fixes a bug in nvtext::edit_distance_matrix where the internal offsets vector is initialized to 0.
This error was introduced in #13912
The bug was found while working on a different PR which re-ordered the nvtext gtests execution causing device memory to be reused from the rmm pool in a different way.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[bdice]

LGTM. Would we want to consider a using ... type declaration, or decltype, or rmm::device_uvector<T>::value_type, or something to reduce the hardcoded types?

[davidwendt]

LGTM. Would we want to consider a using ... type declaration, or decltype, or rmm::device_uvector<T>::value_type, or something to reduce the hardcoded types?

I'm not sure it is worth doing. Every parameter to cudaMemsetAsync is really a potential error.
I came up with this

using offsets_vector_type = rmm::device_uvector<std::ptrdiff_t>;
offsets_vector_type offsets( count, stream );
cudaMemsetAsync( offsets.begin(), 0, count * sizeof(typename offsets_vector_type::value_type), stream.value() );

and realized that I could just have easily put in the wrong count variable too.

[ttnghia]

This looks like something that the compiler can detect and issue a warning :)

[davidwendt]

/merge