Add zeroWhere! fused masked zeroing primitive with ComparisonOp

[Copilot]

The two-pass pattern dz1 *:*= (z1 > 0) allocates a temporary Array[Boolean] and makes two SIMD passes. zeroWhere! fuses compare + blend + store into a single pass with zero allocations.

API

// In-place: zeros vec(i) wherever other(i) satisfies op vs threshold
vec.`zeroWhere!`(other, 0.0f, ComparisonOp.LE)

// Non-mutating: returns new array
val result = vec.zeroWhere(other, 0.0f, ComparisonOp.LE)

// relu backward — one pass, no boolean allocation
dz1.raw.`zeroWhere!`(z1.raw, 0.0f, ComparisonOp.LE)

Changes

• vecxt/src/ComparisonOp.scala — new cross-platform enum ComparisonOp (LT, LE, GT, GE, EQ, NE); exported from all.scala
• JVM (src-jvm/floatarrays.scala, doublearrays.scala) — SIMD implementation using FloatVector/DoubleVector.blend(zero, mask) for the main loop; scalar tail handles non-aligned lengths
• JS/Native (src-js/, src-native/) — while-loop implementations for Array[Float] and Array[Double]
• vecxt/test/src/zeroWhere.test.scala — cross-platform correctness suite: basic correctness, all 6 ops, non-aligned lengths, empty arrays, NaN/infinity edge cases, self-referential arrays; both Float and Double

The op parameter is inline, so the match is eliminated at each call site and the comparison is inlined directly into the loop.

[Quafadas]

@copilot Don’t test the large array in CI , concentrate on correctness only…

[Copilot]

@copilot Don’t test the large array in CI , concentrate on correctness only…

Done — removed the large-array JVM test file in 98f99fa. The remaining tests in vecxt/test/src/zeroWhere.test.scala focus purely on correctness (basic ops, all 6 comparison operators, edge cases, NaN/infinity, etc.).