Delay formatting type errors to improve performance of failed extractions.

[adamreichold]

This should be especially useful for functions that accept several element types like what was discussed in #289. Even more so, as rust-numpy implements PyTypeInfo::is_type_of in terms of FromPyObject::extract.

The benchmark results seem to justify the manual implementation of PyErrArguments:

name             main ns/iter  pr ns/iter  diff ns/iter   diff %   speedup 
extract_failure  13,326        90               -13,236  -99.32%  x 148.07 
extract_success  20            19                    -1   -5.00%    x 1.05 

[adamreichold]

Sadly, this addition of a lifetime parameter to TypeError means this cannot be part of a maintenance release because TypeError is part of the crate API even it is not part of any type signature as it is only ever produced wrapped within PyErr. :-(

[adamreichold]

Decided to actually remove DimensionalityError and TypeError from the API. They are never used directly and cannot be recovered via downcasting either. So the additional API surface does not gain us anything.

[davidhewitt]

Nice speedup! I hadn't looked too hard at optimizing error pathways in PyO3 so far. It looks like we might be introducing some nice speedups into the 0.17 release cycle :)