Graph-mode prepare — auto-wrap the quant boundaries and insert observers.

prepare_fx is the graph-mode entry point of static PTQ, the counterpart to eager lucid.quantization.prepare. It removes the one piece of manual bookkeeping the eager flow demands: instead of hand-placing a QuantStub at the model input and a DeQuantStub at the output, it wraps the model in a lucid.nn.quantized.QuantWrapper so the input is quantized and the output dequantized automatically, then runs the ordinary eager lucid.quantization.prepare underneath to attach activation observers. From there the workflow is identical — run calibration data through the returned model, then convert_fx.

The graph-mode payoff is downstream: because every quantized replacement is an ordinary composite, the converted model traces cleanly through lucid.compile, letting the dispatch-bound quantized graph fuse into a single executable — the regime where quantization's inference win is structurally permanent. (True FX-style observer insertion into a traced graph would need C++ TraceGraph mutation; quantizing at the module level and letting compile capture it reaches the same result.)