fix: avoid variable shadowing of timestep t in compute_loss

[Mr-Neutr0n]

Summary

In cosyvoice/flow/flow_matching.py, the compute_loss() method has a variable shadowing issue where t (the sequence length integer from shape unpacking) is immediately overwritten by t (the random timestep tensor):

b, _, t = mu.shape          # t = sequence length (integer)
t = torch.rand([b, 1, 1], ...)  # t = random timestep (tensor) — immediately overwrites the above

The integer t from the shape unpacking is never used before being reassigned. While this doesn't cause a runtime error in the current code, it is misleading and fragile — a future developer might assume t still holds the sequence length after the unpacking line, leading to subtle bugs.

Fix

Replace b, _, t = mu.shape with b, _, _ = mu.shape to make it explicit that only the batch dimension b is needed from the shape, and that the subsequent t is solely the random timestep tensor.

Test plan

• Verified that t (sequence length) is not referenced anywhere between the unpacking and the reassignment
• Confirmed the change is semantically equivalent — no behavioral change, only improved clarity