lru_caches on the mapper classes are subtly wrong

[ml-evs]

Turns out that lru_cache binds globally where it is defined, so the caches are shared between the different mapper types, e.g., adding a print inside the cached ENDPOINT method of the reference and structure mappers, you see the following behaviour:

>>> from optimade.server.mappers import ReferenceMapper, StructureMapper
>>> ReferenceMapper.ENDPOINT
recomputed for <class 'optimade.server.mappers.references.ReferenceMapper'>
'references'
>>> ReferenceMapper.ENDPOINT
'references'
>>> StructureMapper.ENDPOINT
recomputed for <class 'optimade.server.mappers.structures.StructureMapper'>
'structures'
>>> ReferenceMapper.ENDPOINT
recomputed for <class 'optimade.server.mappers.references.ReferenceMapper'>
'references'

i.e., since the maxsize on the cache is only 1, only successive calls to the same mapper are cached.

This behaviour is a bit unexpected and also be traced to various memory leaks; I don't think we need to worry about these specifically as we are not actually creating instances of the class (see this video for more info).

So, the simplest fix is just to increase the cache sizes so that each of the mapper caches can fit.