fix: use trait-based cross-version compat for parallel collections

Created ParCollectionCompat trait with version-specific implementations:
- Scala 2.12: Empty trait (parallel collections built-in)
- Scala 2.13: Trait with CollectionConverters import

KMeansPlusPlus now extends ParCollectionCompat to get .par support.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

Author: derrickburns

src/main/scala-2.12/com/massivedatascience/clusterer/compat/ParCollectionCompat.scala

@@ -0,0 +1,4 @@
+package com.massivedatascience.clusterer.compat
+
+// Scala 2.12 compatibility: parallel collections are built-in, no import needed
+trait ParCollectionCompat

src/main/scala-2.12/com/massivedatascience/clusterer/compat/package.scala

@@ -0,0 +1,6 @@
+package com.massivedatascience.clusterer.compat
+
+import scala.collection.parallel.CollectionConverters._
+
+// Scala 2.13 compatibility: re-export parallel collection converters
+trait ParCollectionCompat

src/main/scala-2.13/com/massivedatascience/clusterer/compat/ParCollectionCompat.scala

@@ -24,15 +24,14 @@ import com.massivedatascience.util.XORShiftRandom
 
 import scala.annotation.tailrec
 import scala.collection.mutable.ArrayBuffer
-import com.massivedatascience.clusterer.compat._
 
 /** This implements the <a href="http://ilpubs.stanford.edu:8090/778/1/2006-13.pdf">KMeans++ initialization
   * algorithm</a>
   *
   * @param ops
   *   distance function
   */
-class KMeansPlusPlus(ops: BregmanPointOps) extends Serializable with Logging {
+class KMeansPlusPlus(ops: BregmanPointOps) extends Serializable with Logging with compat.ParCollectionCompat {
 
   /** Select centers in rounds. On each round, select 'perRound' centers, with probability of selection equal to the
     * product of the given weights and distance to the closest cluster center of the previous round.