Schema inference reads every file's footer on the driver by default #100

VortexFooterReader.inferSchema merges the schema of every file listed, because vortex.mergeSchema defaults to true (VortexFooterReader.java:92). It reads each footer from the driver, with vortex.footerParallelism threads (default 8, :49). Unlike vortex.stats.maxFiles, nothing caps how many files it reads.

Measured: loading 8,192 files of one directory spent about 6.5 minutes of driver time in mapFooters, about 21 footers a second, while the cluster sat idle. A dataset of about 200k files would take hours before any task starts. Setting vortex.mergeSchema=false brought it down to one footer read.

Proposal, any of:

• Default mergeSchema to false, as Spark's Parquet source does. Callers whose files' schemas differ opt in.
• Or above a threshold of files, merge footers in a Spark job, as Parquet's mergeSchemasInParallel does, rather than on the driver.
• Or cap the files merged as vortex.stats.maxFiles does, and say so in the error if their schemas then disagree.