Improve the implementation of collect() on DataFrame in SparkR

currently in SparkR, collect() on a DataFrame collects the data within the DataFrame into a local data.frame. R users are used to using data.frame.

However, collect() currently can't collect data of nested types from a DataFrame because:
1. The serializer in JVM backend does not support nested types;
2. collect() in R side assumes each column is of simple atomic type that can be combinded into a atomic vector