Optimize memory usage of json reader

[rockyzhengwu]

Is your feature request related to a problem or challenge? Please describe what you are trying to do.

The implementation of decode json to arrow array need convert batch_size of json str to serde_json Value .
this equires a lot of memory for serde_json Value. if with a big batch_size will OOM , usually a large batch_size will have a good compression rate.

arrow-rs/arrow-json/src/reader.rs

Line 685 in e1b5657

let mut rows: Vec<Value> = Vec::with_capacity(batch_size);

Describe the solution you'd like
current implementation in pseudocode:

for batch in value_iter{
   let mut rows: Vec<Value> = Vec::with_capacity(batch_size);
   let arrays = convert_function(rows)
}

If convert ony one json str to serde_json Value will save 3x-5x memory or more, i didn't record carefully .
I had implement a version in our online product in this way , because we use a large batch_size . the pseudocde is

let  field_builder:  Vec<Box<dyn ArrayBuilder>> = create_array_builder(batch_size);
for (i, row) in value_iter.enumerate(){
    let value = serde_json::from_str(row);
    for (index, field) in schema.field.fields{
         let col_name = field.name();
         field_builder[i].append(value.get(col_name))
     }
     if i == batch_size{
        let array_refs = builder.iter_mut().map(|builder| builder.finish()).collect();
        .....
    }
}

this implementation didn't effect the performance.
But it didn't support deep nested list and map.
I'm not sure this is a elegant way for this. or it's possiable to support deep nested list and map.
if this is a good idea , I can try to make a PR for this .

Describe alternatives you've considered

Additional context