The re-introduction of nested item support caused a significant performance degradation

[Gallaecio]

I have a CPU-bound Scrapy project that becomes 50% slower after #29.

I believe the problem is that is_item is not a cheap call, and it will potentially become more expensive as itemadapter extends support to additional types.

I think we may be able to reimplement this function so that it does not call is_item at all, but still does not treat item-like objects as sequences.