BraketAwsQubitDevice unnecessarily computes parameter-shifted gradients for SV1.

[kshyatt-aws]

Describe the bug

Consider a circuit like:

sv1_device_arn = "arn:aws:braket:::device/quantum-simulator/amazon/sv1"
dm1_device_arn = "arn:aws:braket:::device/quantum-simulator/amazon/dm1"
dev = qml.device("braket.aws.qubit", device_arn=sv1_device_arn, wires=nqubits)
nqubits = 4
@qml.qnode(dev)
def quantum_circuit(input, weights):
    # Prepare circuit here
    for i in range(nqubits):
        qml.H(i)
    # Measurement layer
    return [qml.expval(qml.PauliZ(i)) for i in range(nqubits)]

When SV1 is used, this line is triggered. This will compute parameter-shifted derivatives even on the forward pass, when they should not be computed. This makes running with SV1 ~10x slower (and incurs extra costs) compared to DM1.

To reproduce
Run above code snippet.

Expected behavior
The Braket-PL device should only compute the gradient if it's "really" requested.

Screenshots or logs
N/A

System information
A description of your system. Please provide:

• Amazon Braket Python PennyLane Plugin version: 1.23.0
• Amazon Braket Python SDK version: 1.66.0
• Amazon Braket Python Schemas version: N/A
• Amazon Braket Python Default Simulator version: N/A
• Python version: 3.9

Additional context
One other solution would be to avoid calling execute_and_gradients entirely in this case.

[speller26]

The problem is execute_and_gradients being automatically called if the device supports its own gradient computation method and the execution is analytic (shots = 0); this is decided upstream by PennyLane, and isn't controlled in the plugin. However, I've added a warning to the parameter-shift fallback path.

[speller26]

Did not intend to auto-close the issue

[laurencap]

@kshyatt-aws confirmed with me that we can close the issue, since any additional work would be outside this repo. Thanks!