Widen tolerance for random tests on WASI

With only 1K elements (vs 1M native), the statistical variance is much
higher. The tolerance bounds are now adjusted based on sample size:
- WASI: ±0.10 for rand, ±0.50 for randint (1K samples)
- Native: ±0.02 for both (1M samples)

Author: pedrovgs

Tests/MatftTests/RandomTest.swift

@@ -13,19 +13,22 @@ final class RandomTests: XCTestCase {
     func testRand() {
         // Use smaller array size on WASM to avoid memory pressure
         // 100x100x100 = 1M elements = 4MB per array, too much for WASM
+        // Smaller sample size requires wider tolerance for mean (higher variance)
         #if os(WASI)
         let shape = [10, 10, 10]  // 1K elements
+        let tolerance: Float = 0.10  // Wider bounds for smaller sample
         #else
         let shape = [100, 100, 100]  // 1M elements
+        let tolerance: Float = 0.02  // Tight bounds for large sample
         #endif
 
         let a = Matft.random.rand(shape: shape)
         let mean_a = a.mean().scalar as! Float
-        let cond_a = (Float(0.48) < mean_a) && (mean_a < Float(0.52))
+        let cond_a = (Float(0.5 - tolerance) < mean_a) && (mean_a < Float(0.5 + tolerance))
 
         let b = Matft.random.rand(shape: shape)
         let mean_b = b.mean().scalar as! Float
-        let cond_b = (Float(0.48) < mean_b) && (mean_b < Float(0.52))
+        let cond_b = (Float(0.5 - tolerance) < mean_b) && (mean_b < Float(0.5 + tolerance))
 
         XCTAssertTrue(cond_a, "The result may be failure due to random generation")
         XCTAssertTrue(cond_b, "The result may be failure due to random generation")
@@ -35,19 +38,22 @@ final class RandomTests: XCTestCase {
 
     func testRandint() {
         // Use smaller array size on WASM to avoid memory pressure
+        // Smaller sample size requires wider tolerance for mean (higher variance)
         #if os(WASI)
         let shape = [10, 10, 10]  // 1K elements
+        let tolerance: Float = 0.50  // Wider bounds for smaller sample
         #else
         let shape = [100, 100, 100]  // 1M elements
+        let tolerance: Float = 0.02  // Tight bounds for large sample
         #endif
 
         let a = Matft.random.randint(low: 0, high: 10, shape: shape)
         let mean_a = a.mean().scalar as! Float
-        let cond_a = (Float(4.48) < mean_a) && (mean_a < Float(4.52))
+        let cond_a = (Float(4.5 - tolerance) < mean_a) && (mean_a < Float(4.5 + tolerance))
 
         let b = Matft.random.randint(low: 0, high: 10, shape: shape)
         let mean_b = b.mean().scalar as! Float
-        let cond_b = (Float(4.48) < mean_b) && (mean_b < Float(4.52))
+        let cond_b = (Float(4.5 - tolerance) < mean_b) && (mean_b < Float(4.5 + tolerance))
 
         XCTAssertTrue(cond_a, "The result may be failure due to random generation")
         XCTAssertTrue(cond_b, "The result may be failure due to random generation")