Fix batch size mismatch in clifford_qnn example

[yurekami]

Summary

• Add reset_states(bsz) call in QFCModel.forward() to ensure quantum device state dimensions match the current batch size
• The QuantumDevice is initialized with bsz=1 by default, but during training the batch size varies
• Without resetting states, tensor dimension mismatch causes RuntimeError during matrix operations

Root Cause

The QuantumDevice is created in __init__ with default batch size 1:

self.q_device = tq.QuantumDevice(n_wires=self.n_wires)  # bsz defaults to 1

But in forward(), the input batch size varies (256 for full batches, smaller for final batch). The quantum operations expect the device states to match the input batch size.

Fix

Call reset_states(bsz) before using the quantum device:

def forward(self, x, use_qiskit=False):
    bsz = x.shape[0]
    x = F.avg_pool2d(x, 6).view(bsz, 16)
    
    # Reset quantum device states for current batch size
    self.q_device.reset_states(bsz)
    
    # ... rest of forward

Test plan

• Run python examples/clifford_qnn/mnist_clifford_qnn.py to verify no RuntimeError
• Verify training completes with varying batch sizes

Fixes #213

🤖 Generated with Claude Code