Copy the input in thresholded_relu's value_inference

[LeSingh1]

Problem

thresholded_relu.value_inference aliases its input's array and then writes through it:

@precondition(allow=VALUE)
def value_inference(self):
    y = self.x.val          # alias, not a copy
    y[y < self.alpha.val] = 0
    return y

Every sibling op in the same file copies first — elu (line 107), leaky_relu (line 210), softplus_parametric (lines 497-498) all start with np.copy(...). thresholded_relu does not, so constant-folding it overwrites the value of its input const.

When that const is also used elsewhere — as a weight, say — the weight is silently replaced with the thresholded values:

w_val = np.array([[-1.0, 0.5], [2.0, -3.0]], dtype=np.float32)

@mb.program(input_specs=[mb.TensorSpec(shape=(2, 2))])
def prog(x):
    w = mb.const(val=w_val, name="weight")
    y = mb.matmul(x=x, y=w)
    z = mb.thresholded_relu(x=w, alpha=1.0)
    return mb.add(x=y, y=z)

	value
matmul's weight operand	[[0.0, 0.0], [2.0, 0.0]]
correct	[[-1.0, 0.5], [2.0, -3.0]]

The corrupted array is what gets serialized into the weight blob. The caller's own numpy array is clobbered as well, since the const does not copy it on the way in.

Fix

y = np.copy(self.x.val), matching the other activation ops in the file.

Tests

In TestThresholdedReLU (ops/tests/iOS14/test_activation.py):

• test_builder_eval_does_not_mutate_input — the input array must be unchanged after folding, and the folded value must still be correct.
• test_builder_eval_does_not_corrupt_shared_const — the shared-weight program above; matmul's operand must keep its original value.

Both fail on main and pass with the fix.

The existing test_builder_eval is untouched and still passes. It happens to compute its expected value as y = x_val; y[y < 2.0] = 0 on the same array value_inference was mutating, which is why it never caught this.

I ran all of ops/tests/iOS14/test_activation.py before and after; the pre-existing failure set is identical, 188 either way (this environment cannot load CoreML.framework, so the run_compare_builder tests fail there regardless).