What are the libararies for Machine Learning on NISQ Chip? And What are the roadmaps?

You will have very limited support simulating photonic circuits in Cirq (and therefore Tensorflow Quantum). Those libraries mainly deal with evolving systems of qubits where a state $|psirangle$ is expressed over a discrete basis like $|psirangle = sum_i c_i |irangle$, as opposed to photonic systems containing continous variables (such as a wavefunction $psi(x)$ expressed in the basis of a continous operator $hat{x}$).

One of the main uses for Tensorflow Quantum is when you have an algorithm that involves both a quantum simulation part and a classical processing part that interact. For example a parameterized quantum circuit and some kind of neural network layer that are both trained using gradient-based optimizers, then the library is able to efficiently simulate both the circuit and its derivatives which are then propagated through both of the layers to differentiate a given cost function. This does not sound like a good fit for your application since you would need to outsource the circuit simulation to a different program.

Maybe check out Strawberry Fields since this was designed for photonic simulation and interfaces with Xanadu's Pennylane and so should have similar capabilities as I described above, but for a photonic qc simulator. I'm not very familiar with the library so I don't know what kind of support it has for device noise.