perf: express geometric residual as one addcmul operator

[DivyamTalwar]

Summary

• replace x + delta * k with torch.addcmul(x, k, delta)
• retain the literal implementation as an auditable reference
• test forward and backward parity
• add a local latency/CUDA peak-allocation benchmark

Acceptance

• default and reference output/gradients agree
• benchmark reports measured results rather than claiming a universal speedup
• CI must pass before merge