Fix pipelined sampling corruption: use per-call execution descriptor

The MPSGraphCompositeSampler reused a single
MPSGraphExecutableExecutionDescriptor across all pipelined steps.
Under pipelined execution (depth > 1), overlapping runAsync calls
on the same executable corrupt intermediate scratch buffers when
sharing a descriptor, producing garbled output (word repetitions,
doubled punctuation) with temperature > 0.

Create a fresh descriptor per encode() call, matching the pattern
MPSGraphArgmaxSampler already uses.

Author: sukru tikves

swift/Sources/CoreAILanguageModels/Samplers/MPSGraphSamplers.swift

@@ -496,7 +496,6 @@ final class MPSGraphCompositeSampler: @unchecked Sendable {
     private let randomData: MPSGraphTensorData
     private let topPData: MPSGraphTensorData
     private let minPData: MPSGraphTensorData
-    private let execDescriptor: MPSGraphExecutableExecutionDescriptor
 
     /// Testing only: Override random value for deterministic tests.
     var testingOnlyRandomOverride: Float?
@@ -678,7 +677,6 @@ final class MPSGraphCompositeSampler: @unchecked Sendable {
             shape: [1 as NSNumber],
             dataType: .float32
         )
-        self.execDescriptor = MPSGraphExecutableExecutionDescriptor()
     }
 
     /// Encode composite sampling asynchronously (protocol conformance).
@@ -725,7 +723,10 @@ final class MPSGraphCompositeSampler: @unchecked Sendable {
             cachedOutputBuffer = outputBuffer
         }
 
-        execDescriptor.completionHandler = { [outputBuffer, outputOffset] (_, error) in
+        // Per-call descriptor — reusing one across pipelined steps corrupts
+        // intermediate scratch buffers when multiple runAsync calls overlap.
+        let desc = MPSGraphExecutableExecutionDescriptor()
+        desc.completionHandler = { [outputBuffer, outputOffset] (_, error) in
             if let error = error {
                 print("MPSGraph composite sampler error: \(error)")
                 completion(0)
@@ -743,7 +744,7 @@ final class MPSGraphCompositeSampler: @unchecked Sendable {
             with: queue,
             inputs: [logitsData, temperatureData, randomData, topPData, minPData],
             results: [outputData],
-            executionDescriptor: execDescriptor
+            executionDescriptor: desc
         )
     }