ENH: Implement use_nullable_dtypes properly for read_json

[phofl]

The option was added in #50750 to be ready for 2.0. We should do the dtype inference before actually creating the DataFrame/Series, this should improve performance

[kostyafarber]

Would be interesting in working on this. What would be a good start?

[phofl]

The main idea is the following: We create numpy-arrays before converting them to a DataFrame. Right now we are converting to extension arrays after creating the DataFrame. Ideally, we want to convert to Extension Arrays before even creating the DataFrame

[kostyafarber]

Will this need tests? or are the tests you did in #50750 enough?

[phofl]

Probably enough except if you discover things that are not covered