Do not decompose non-trainable operations in adjoint

[astralcai]

Context:
The following code throws an error trying to decompose non-trainable QubitUnitary.

import pennylane as qml
from pennylane import numpy as pnp

@qml.qnode(qml.device("default.qubit", wires=3), diff_method="adjoint")
def circuit(x):
    qml.RX(x, 0)
    qml.QubitUnitary(np.eye(8), [0, 1, 2])
    return qml.expval(qml.PauliZ(2))

x = pnp.array(1.1, requires_grad=True)
qml.jacobian(circuit)(x)

Description of the Change:
Add qml.operation.is_trainable check to the stopping condition for adjoint ops.

Benefits:
Bug Fix

Related GitHub Issues:
#5217

Related Shortcut Stories
[sc-56905]

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (a8be574) 99.68% compared to head (a2296f7) 99.68%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #5233      +/-   ##
==========================================
- Coverage   99.68%   99.68%   -0.01%     
==========================================
  Files         399      399              
  Lines       36793    36515     -278     
==========================================
- Hits        36678    36399     -279     
- Misses        115      116       +1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[albi3ro]

Thanks for fixing this 🎉