From SERVER-119549, we add a cursor overhead of 2KB for the estimated per-iterator (per-cursor) memory cost used during the merge phase of the external sort.

We want to:
1. Enumerate the components currently assumed to play a part in the overhead:

• WT-internal cursor allocation (~800 B) — verify against __wt_cursor + btree cursor structs in src/third_party/wiredtiger/src/include/cursor.h and btree.h
• Mongo SeekableRecordCursor / WT cursor wrapper (~70 B) — sizeof the concrete class in wiredtiger_record_store_cursor.h.
• MergeIterator::Stream + copied Key (~40 B + Key) — sizeof(Stream) plus the serialized Key buffer.
• WT-internal string/key buffers (variable) — not easy to sizeof; needs runtime measurement.

2. (easier pass than 3.) Measure the overhead across different platforms using sizeof()
3. Use a memory profiler to measure the overhead 
4. Vary inputs of key sizes
5. Update the cursor overhead accordingly.