[executorch][llm] Make Gemma Apple test deterministic

Pull Request resolved: pytorch#21403

`MultimodalRunnerTest.testGemma` asserted on generated text while inheriting the default sampling temperature of 0.8. Set the temperature to zero for both generation calls so all four TestInfra configurations use deterministic greedy decoding. This commit was authored with Claude.
ghstack-source-id: 407467872
@exported-using-ghexport

Differential Revision: [D113786993](https://our.internmc.facebook.com/intern/diff/D113786993/)

Author: SS-JIA

extension/llm/apple/ExecuTorchLLM/__tests__/MultimodalRunnerTest.swift

@@ -123,6 +123,7 @@ class MultimodalRunnerTest: XCTestCase {
         MultimodalInput(String(format: chatTemplate, userPrompt)),
       ], Config {
         $0.sequenceLength = sequenceLength
+        $0.temperature = 0
       }) { token in
         text += token
         if token == "<end_of_turn>" {
@@ -143,6 +144,7 @@ class MultimodalRunnerTest: XCTestCase {
         MultimodalInput(String(format: chatTemplate, userPrompt)),
       ], Config {
         $0.sequenceLength = sequenceLength
+        $0.temperature = 0
       }) { token in
         text += token
         if token == "<end_of_turn>" {