[FEA][JNI] Consider defaulting parquet dictionary encoding policy to ALWAYS

[abellina]

This PR is going to set the cuDF dictionary encoding policy for parquet to ADAPTIVE (#15570)

This is to get around an issue in nvcomp zstd #15501, where too large pages are getting created and is causing zstd to not compress larger dictionary pages.

For now we can pick ALWAYS in order to retain the current behavior for Spark. We should consider the impact of this setting for different compression, especially zstd (reading and writing).