Revert regression from PR elastic#98139 affecting model deployment up…

…dates

This commit reverts changes to the memory usage estimation logic introduced by PR elastic#98139, which caused failures when updating the `number_of_allocations` for trained model deployments. The reversion restores the system's stability in high-availability environments.

Relates elastic#107807

x-pack/plugin/core/src/main/java/org/elasticsearch/xpack/core/ml/action/StartTrainedModelDeploymentAction.java

@@ -714,26 +714,7 @@ public static long estimateMemoryUsageBytes(
         int numberOfAllocations
     ) {
         // While loading the model in the process we need twice the model size.
-
-        // 1. If ELSER v1 or v2 then 2004MB
-        // 2. If static memory and dynamic memory are not set then 240MB + 2 * model size
-        // 3. Else static memory + dynamic memory * allocations + model size
-
-        // The model size is still added in option 3 to account for the temporary requirement to hold the zip file in memory
-        // in `pytorch_inference`.
-        if (isElserV1Or2Model(modelId)) {
-            return ELSER_1_OR_2_MEMORY_USAGE.getBytes();
-        } else {
-            long baseSize = MEMORY_OVERHEAD.getBytes() + 2 * totalDefinitionLength;
-            if (perDeploymentMemoryBytes == 0 && perAllocationMemoryBytes == 0) {
-                return baseSize;
-            } else {
-                return Math.max(
-                    baseSize,
-                    perDeploymentMemoryBytes + perAllocationMemoryBytes * numberOfAllocations + totalDefinitionLength
-                );
-            }
-        }
+        return isElserV1Or2Model(modelId) ? ELSER_1_OR_2_MEMORY_USAGE.getBytes() : MEMORY_OVERHEAD.getBytes() + 2 * totalDefinitionLength;
     }
 
     private static boolean isElserV1Or2Model(String modelId) {