change json vector reader to do direct fuse

[mccanne]

This commit changes the json vector reader to create an fused output with fusion types directly instead of using unblend to create a Dynamic. These changes get at the fundamental bottleneck of computing types over the variant/json pattern.

To implement this, we added a couple methods to TypeDefs to be able to enter array and record types directly without having to have a corresponding super.Type. We also created TypeDefsMapper to efficiently create super.Types from a TypeDefs table and cache intermediary types. This is used by the loader on the vector.Fusion that is created by the JSON reader to create the subtype super.Types on demand when needed.

In particular, the CSUP writer currently invokes the loader to create the super.Types then converts them to typevals that are stored in the subtypes table in the CSUP metadata. The next step to improve this write path is the store the TypeDefs in the CSUP metadata instead of type values and never materialize the subtype IDs to types if they are never otherwise needed. This means large amounts of JSON can be read, processed, and written into CSUP and the query type context does not need to be polluted with all the fusion subtypes.

We also put an option into samFunc call to avoid ripping vectors when fusion vectors are passed to defuse.

Finally we increased the size of fjson batch size from 1K to 10K.

[nwt]

Is this based on measurement or intuition? If intuition, consider that DuckDB's default is 2048.

[mccanne]

Let's merge and address again after measurement.

[nwt]

This is fine for now but we do have an existing pattern for per-function behaviors like this (see the RipUnions interface).