Unify CPU and GPU tree transformer code paths #578

Follow-up to #556 (see #556 (comment)).

The GPU path already passes the flat data vectors to add_transform_kernel! and keeps all the subblock structure (strides/offsets) inside the device transformer. The CPU path instead builds StridedSubblocks for both tensors and passes those in alongside the transformer. Doing the same on the CPU, i.e. passing the flat data vector and keeping the subblock information in the transformer, would make the two code paths more uniform and easier to follow.

This changes the CPU kernel signatures, so it should be benchmarked against the current implementation.

🤖 Generated with Claude Code