Add custom decoder in arrow-json

[cht42]

Similar behavior as #7015 but on the decoder side. This would allow introducing flexibility when decoding JSON.

For example, when trying to replicate JSON decoding in spark, spark will decode incorrect values as null or raw string. This would allow overriding the current StringDecoder to implement this functionality.

[Blizzara]

FYI @tustvold @alamb @adriangb since you worked on the encoder one. This would allow us (I work with @cht42) to not vendor the code to get the spark-compatibility. Alternatively, if you have other ideas for how to reach our goal (we could also make the spark compat a config option, for example), let us know!

[tustvold]

I'm afraid I am unlikely to have time to review this in the near future, but I am not a big fan of making the tape public, this can and has changed in non-trivial ways in the past.

get the spark-compatibility

What does this look like? Is it just using nulls for fallible conversions, as that doesn't sound like a big lift, or is it something more intrusive?

[cht42]

It is more intrusive when the target type is string. In that case it will use the raw string instead of null. For example

{"a": [1,2,3]}

With target schema of map<str, str> will be decoded as

{"a": "[1,2,3]"}

In spark.

For just fallback to null on parsing errors, #7230 is open that would handle that.

[adriangb]

Without having looked at the code I will try to review - although I may not be able to give much feedback since I haven't worked on this side of the codebase before and we do not use it internally.