Fix Apple Silicon MPS compatibility and refactor native dtype/device casting

[sansiro77]

Description

This PR resolves the Apple Silicon MPS backend issues reported in #149 , and significantly refactors the internal device/dtype casting mechanism to align with PyTorch's native behaviors.

Key Changes

1. Fix Apple Silicon MPS Compatibility

• Resolve Parametric Gates Stacking: Explicitly cast tensors to complex dtype before torch.stack in parametric gates (e.g., Rx). This fixes the cat_int32_t_float_float2 RuntimeError on the MPS backend.
• Auto-fallback to CPU for >16 dimensions: Add a guard in the forward pass. If the state tensor exceeds the 16-dimension limit, it throws a clear warning and automatically falls back to CPU.

2. Refactor PyTorch Native Casting

• Replace Custom .to() with ._apply(): Transition all custom nn.Modules from overriding .to() to using ._apply() to enable robust, native recursive graph traversal.
• Implement Unified Complex Probing: Add apply_complex_fix inside ._apply() to safely map floating-point actions to their complex counterparts during device/dtype transitions.

Limitations & Known Issues

• Gaussian Backend MPS Support: Due to current limitations in PyTorch's MPS backend, torch.linalg.det() for complex-valued matrices is not yet supported, resulting in RuntimeError: linalg.lu_factor(): MPS doesn't support complex types.

[Hugh-888]

Test examples:

• Fock backend with basis=True

print("torch", torch.__version__)
print("mps available", torch.backends.mps.is_available())

nmode=4
cir = dq.QumodeCircuit(nmode=4, init_state='vac', backend='fock', basis=True, cutoff=2)
for i in range(nmode-1):
    cir.bs([i, i+1])
cir.homodyne(0)
cir.to("mps")
prob = cir(is_prob=True)

• Fock backend with basis=False

print("torch", torch.__version__)
print("mps available", torch.backends.mps.is_available())

nmode=4
cir = dq.QumodeCircuit(nmode=4, init_state='vac', backend='fock', basis=False, cutoff=2)
for i in range(nmode-1):
    cir.bs([i, i+1])
cir.homodyne(0)
cir.to("mps")
prob = cir(is_prob=True)

• Homodyne measurement for Fock backend with basis=False

print("torch", torch.__version__)
print("mps available", torch.backends.mps.is_available())

nmode=4
cir = dq.QumodeCircuit(nmode=4, init_state='vac', backend='fock', basis=False, cutoff=2)
for i in range(nmode-1):
    cir.bs([i, i+1])
cir.homodyne(0)
cir.to("mps")
state = cir()
samples = cir.measure_homodyne()

• Gaussian backend
  Due to current limitations in PyTorch's MPS backend, torch.linalg.det() for complex-valued matrices is not yet supported, resulting in RuntimeError: linalg.lu_factor(): MPS doesn't support complex types.

print("torch", torch.__version__)
print("mps available", torch.backends.mps.is_available())

nmode=4
cir = dq.QumodeCircuit(nmode=4, init_state='vac', backend='Gaussian', basis=True, cutoff=2)
for i in range(nmode-1):
    cir.bs([i, i+1])
cir.homodyne(0)
cir.to("mps")
prob = cir(is_prob=True)