braket.aws.qubit returns incorrect expval from a ProgramSetQuantumTaskResult (measurements are correct)

### Summary
When a `qml.expval` QNode runs on `braket.aws.qubit` against a backend that returns a **`ProgramSetQuantumTaskResult`** (observed on Rigetti Cepheus-1-108Q, where the plugin submits a parametric program set), the returned expectation values are wrong. The underlying measurement bitstrings are **correct**; only the plugin's computed expectation is wrong, and a `UserWarning: No observable was measured` is emitted.

### Steps to reproduce
```python
import numpy as np, pennylane as qml
ARN = "arn:aws:braket:us-west-1::device/qpu/rigetti/Cepheus-1-108Q"
dev = qml.device("braket.aws.qubit", device_arn=ARN, wires=3, shots=200)

@qml.qnode(dev)
def circuit(t):
for i in range(3):
qml.RY(t[i], wires=i)
return [qml.expval(qml.PauliZ(i)) for i in range(3)]

print(circuit(np.array([0.4, 1.2, 2.0])))
```
A minimal parametric circuit is enough to trigger the program set on a QPU with parametric compilation. **SV1 does not reproduce** (it returns a `GateModelQuantumTaskResult` with correct expvals), so a QPU is required.

### Observed (Cepheus, 3 qubits, 200 shots, `t = [0.4, 1.2, 2.0]`)
| | wire 0 | wire 1 | wire 2 |
|---|---:|---:|---:|
| **plugin `expval`** | 0.77 | **0.77** | **0.77** |
| **correct** (sample mean `1 − 2·mean(bit)` from the same result's measurements) | 0.77 | 0.41 | −0.25 |

The plugin returns **the first observable's value for all observables** three different angles all report 0.77, though the true per-qubit expectations differ. On a 7-qubit workload the divergence reaches ~1.5.

### Expected
For diagonal Pauli-Z observables, `expval` should equal the sample mean `1 − 2·mean(bit)` per qubit, recoverable directly from `result.entries[0].entries[0].measurements`.

### Root cause
For program-set-capable devices the plugin submits **without attaching observables**, in `src/braket/pennylane_plugin/braket_device.py` (present on current `main`):
```python
add_observables=not self._supports_program_sets,
```
So the program-set entries come back with no observable attached, and `amazon-braket-sdk`'s `MeasuredEntry.expectation()` then returns `None` (unimplemented `# TODO`, see amazon-braket/amazon-braket-sdk-python#1316); the plugin produces an incorrect expectation instead of computing the sample mean from `measurements`.

### Workaround
```python
res = dev._task.result() # or recover the task; ProgramSetQuantumTaskResult
meas = np.asarray(res.entries[0].entries[0].measurements) # shots x wires
z = 1.0 - 2.0 * meas.mean(axis=0) # correct diagonal expectations
```

### Suggested fix
When a program-set entry lacks an attached observable, compute the requested expectation from the entry's `measurements` rather than trusting `MeasuredEntry.expectation()`.

### Environment
- `pennylane==0.44.1`
- `amazon-braket-pennylane-plugin==1.34.0` — **reproduced on this version.** Not re-run on 1.34.2, but the responsible code path (`add_observables=not self._supports_program_sets`) is present on current `main`, and the 1.34.1 / 1.34.2 changelogs show no expectation-handling fix.
- `amazon-braket-sdk==1.113.1` — the SDK `# TODO` is also still present on `main` (1.121.0).
- Device: Rigetti Cepheus-1-108Q.