Auto merge of #82162 - cuviper:flat-fold, r=Mark-Simulacrum

Auto merge of #82162 - cuviper:flat-fold, r=Mark-Simulacrum

Expand FlattenCompat folds

The former `chain`+`chain`+`fold` implementation looked nice from a
functional-programming perspective, but it introduced unnecessary layers
of abstraction on every `flat_map`/`flatten` fold. It's straightforward
to just fold each part in turn, and this makes it look like a simplified
version of the existing `try_fold` implementation.

For the `iter::bench_flat_map*` benchmarks, I get a large improvement in
`bench_flat_map_chain_sum`, from 1,598,473 ns/iter to 499,889 ns/iter,
and the rest are unchanged.