Qnode primitive captures unprocessed ExecutionConfig (#8258)

**Context:**

Blocks PennyLaneAI/catalyst#2041

**Description of the Change:**

The capture `Device.setup_execution_config` was about the capabilities
of the capture execution pipeline, which is no longer maintained or
actively developed. While I'm not sure where will be the best place for
this logic in catalyst, we can leave that up to catalyst and no
determine it in program capture.

**Benefits:**

Unblocks catalyst unification work.

**Possible Drawbacks:**

**Related GitHub Issues:**

[sc-99193]

---------

Co-authored-by: Andrija Paurevic <[email protected]>

Author: albi3ro

doc/releases/changelog-dev.md

@@ -121,6 +121,10 @@
 
 <h3>Improvements 🛠</h3>
 
+* The `QNode` primitive in the experimental program capture now captures the unprocessed `ExecutionConfig`, instead of
+  one processed by the device.
+  [(#8258)](https://github.com/PennyLaneAI/pennylane/pull/8258)
+
 * The function :func:`qml.clifford_t_decomposition` with `method="gridsynth"` are now compatible 
   with quantum just-in-time compilation via the `@qml.qjit` decorator.
   [(#7711)](https://github.com/PennyLaneAI/pennylane/pull/7711)

pennylane/workflow/_capture_qnode.py

@@ -163,6 +163,8 @@ def _get_shapes_for(*measurements, shots=None, num_device_wires=0, batch_shape=(
 @qnode_prim.def_impl
 def _(*args, qnode, device, execution_config, qfunc_jaxpr, n_consts, shots_len, batch_dims=None):
 
+    execution_config = device.setup_execution_config(execution_config)
+
     if shots_len == 0:
         shots = None
         non_shots_args = args
@@ -354,7 +356,7 @@ def _finite_diff(args, tangents, **impl_kwargs):
 
 @debug_logger
 def _qnode_jvp(args, tangents, *, execution_config, device, qfunc_jaxpr, **impl_kwargs):
-
+    execution_config = device.setup_execution_config(execution_config)
     if execution_config.use_device_gradient:
         return device.jaxpr_jvp(qfunc_jaxpr, args, tangents, execution_config=execution_config)
 
@@ -567,7 +569,6 @@ def f(x):
         config = construct_execution_config(
             qnode, resolve=False
         )()  # no need for args and kwargs as not resolving
-        config = qnode.device.setup_execution_config(config)
 
         if abstracted_axes:
             # We unflatten the ``abstracted_axes`` here to be have the same pytree structure

tests/capture/workflow/test_capture_qnode.py

@@ -101,14 +101,14 @@ def circuit(x):
     assert eqn0.params["qnode"] == circuit
     assert eqn0.params["shots_len"] == 0
     expected_config = qml.devices.ExecutionConfig(
-        gradient_method="backprop",
-        use_device_gradient=True,
+        gradient_method="best",
+        use_device_gradient=None,
         gradient_keyword_arguments={},
         use_device_jacobian_product=False,
         interface="jax",
         grad_on_execution=False,
-        device_options={"max_workers": None, "rng": dev._rng, "prng_key": None},
-        mcm_config=qml.devices.MCMConfig(mcm_method="deferred", postselect_mode=None),
+        device_options={},
+        mcm_config=qml.devices.MCMConfig(mcm_method=None, postselect_mode=None),
     )
     assert eqn0.params["execution_config"] == expected_config
 
@@ -231,15 +231,15 @@ def circuit():
     assert jaxpr.eqns[0].primitive == qnode_prim
     expected_config = qml.devices.ExecutionConfig(
         gradient_method="parameter-shift",
-        use_device_gradient=False,
+        use_device_gradient=None,
         grad_on_execution=False,
         derivative_order=2,
         use_device_jacobian_product=False,
         mcm_config=qml.devices.MCMConfig(
             mcm_method="single-branch-statistics", postselect_mode=None
         ),
         interface=qml.math.Interface.JAX,
-        device_options={"max_workers": None, "rng": dev._rng, "prng_key": None},
+        device_options={},
     )
     assert jaxpr.eqns[0].params["execution_config"] == expected_config