[Relax] Eager free original weights in transform_params

[vinx13]

This updates LiftTransformParams to always free original weight in the bundled weights tuple. This prevents the issue that transform_params consumes memory with double weight size. See discussion https://discuss.tvm.apache.org/t/inplacetransformparams/16323.

cc @tqchen @MasterJH5574 @Lunderberg

[Lunderberg]

Any idea why the KillAfterLastUse pass isn’t already handling this case? It seems like it shouldn’t require special handling in LiftTransformParams.

[tqchen]

@Lunderberg in this case we are also killing the elements within tuple which is not covered by killafter use

[Lunderberg]

Ah, thank you, and that makes sense. That would also explain why the increased footprint isn’t seen when using LazyTransformParams, since the tuple doesn’t exist as a parameter.

If I understand correctly, we could achieve a similar effect by unpacking the entire tuple on its first use, though that could be less robust to later lowering passes. Since we have the in-place updates, it makes sense to use them, and I like the change.