Correct coefficient propagation in the operators

[Cryoris]

Summary

The coefficients of some operators (especially SummedOp and ComposedOp) are lost on conversion, which is fixed. Also state coefficients in an expectation should be applied squared, since the state is applied from both sides.

Details and comments

Some before-and-after comparisons:

op = 0 * (I + Z)
state = Plus
print((~StateFn(op) @ state).eval())
# before: (1+6.1e-17j), coefficient ignored!
# now: 0j

op = I
state = 0 * (Plus + Minus)
backend = Aer.get_backend('qasm_simulator')
sampler = CircuitSampler(backend).convert(~StateFn(op) @ state)
print(sampler.eval()) 
# before: (2+0j), coefficent ignored!
# now: 0j

op = I
state = (Plus + Minus) / numpy.sqrt(2)
sampler = CircuitSampler(backend).convert(~StateFn(op) @ state)
print(sampler.eval())
# before: (2+0j), same as with 0 coefficient
# now: (1+0j), correctly normalized