Fix BulkSetAmpl bug, Add more diff cuquantum unit tests

[jaeyoo]

From this PR, other differentiators like ParameterShift(), LinearCombination and its subclasses can be accelerated with cuquantum expectation and sampled_expectation op.

[jaeyoo]

I also fixed the random seeds at the end of each test in expectation_test and sampled_expectation_test because they are using internal keras layer with randomly initialized weights.

I also found and fixed that the two keras layers were using predefined initializer at the function argument default value position of __init__(). In python, this creates an object and cached it, and then re-use it whenever the new object is created. That doesn't make sense, so Keras team also guided to avoid the usage. After moving it into function body, we can see the errors from the same random initializer gone.

Also, I added initializers in each test in expectation_test and sampled_expectation_test to fix the random behavior in the test.

Now I confirmed that both ./scripts/test_all.sh gpu and ./scripts/test_all.sh passed.