RDKitConverter improvements

[cbouy]

Running the RDKitConverter is slow and its accuracy could always be improved.

Here are a few ideas for improvements:

Category	Description	Status
Speed	Use RDKit's RunReactantInPlace instead of using RunReactants while transferring atom properties from the reactant mol to the product which causes a bottleneck.	Merged in PR #4082
Both	Sorting atoms by their number of heavy atom neighbors (ascending order) before running the inferring code should help in not having to patch conjugated systems in some cases.	Merged in PR #4082
Both	This line should probably exclude this case: [O-]-C(=O)
Accuracy	Analyse the failing molecules from the RDKitConverterBenchmark to identify patterns and handle them with bespoke reactions. Related issue: #3339
Both	RDKit 2022.09.1 added a C++ XYZParser which also includes code to infer bond orders and charges from the topology which could be reused here as an alternative implementation. The only downside is that it requires to know in advance the total charge of the molecule. But if the total charge is known, it should be much faster than our current code. Also adding a simple wrapper around AssignBondOrdersFromTemplate would be very useful.	In progress in PR #4305