PERF: regression in DataFrame construction from nested dict

[jorisvandenbossche]

#41785 removed some cython code to handle this case (because it could also be handled by existing python code, AFAIU), but this de-duplication caused a big slowdown in one of the benchmarks: https://pandas.pydata.org/speed/pandas/#frame_ctor.FromDicts.time_nested_dict_int64?python=3.8&Cython=0.29.21

It's of course a trade-off between the maintenance cost of having the cython version versus the performance benefit. But if the benchmark is representative, a 5-6x slowdown seems quite a lot for getting rid of a relatively small piece of cython code.

Originally posted by @jorisvandenbossche in #41785 (comment)

[jbrockmendel]

can you verify this?

This appears correct. For context, this change was made as a follow-up to a bugfix where this de-duplication would have prevented the bug. I agree 5-6x is a bigger slowdown than I expected.