DataFrameReader doesn't respect MERGE_SCHEMA setting for Parquet

In 1.4.0, parquet is read by DataFrameReader.parquet, when creating ParquetRelation2 object, "parameters" is hard-coded as "Map.empty[String, String]", so ParquetRelation2.shouldMergeSchemas is always true (the default value).
In previous version, spark.sql.hive.convertMetastoreParquet.mergeSchema config is respected.
This bug downgrade performance a lot for a folder with hundreds of parquet files and we don't want a schema merge.