Bug Description

   Problem: When using parallel batch processing (parallelism="threads" or
   parallelism="processes"), PyMongoArrow fails when different batches have different schemas due
   to type inference.

   Specific scenario:
     • First batch contains small integers (inferred as int32)
     • Later batch contains large integers requiring int64
     • The parallel code paths process each batch independently, creating separate Arrow tables
     • When concatenating these tables with pa.concat_tables(), the schemas don't match (int32 vs
       int64), causing an error

   Root causes:
     1. In `api.py`: Used promote_options="default" which requires exact schema matches and doesn't
        allow type promotion
     2. In `lib.pyx`: When promoting int32→int64 during schema inference, the old builder was
        discarded, losing all previously appended int32 values

   The fix:
     1. Change promote_options="permissive" to allow type promotion when concatenating tables
     2. Preserve existing int32 values by casting them to int64 and re-appending to the new int64
        builder

   This ensures parallel and non-parallel code paths produce consistent results when schema
   inference encounters mixed integer sizes.