Improve generated CUDA code readability by renaming `buf_shmem` on demand

`buf_shmem` accesses can be split into two steps: one giving buffer base address a name like `float *x_smem = buf_shmem + offset`, the other being real access but now with `x_smem[...]`.

This should not hinder shared memory reuse since during the lifetime of one buffer it's never moved around, and multiple pointers can share overlapped storage.

The issue does not name a file, test, or entry point. Start by locating the CUDA code-generation path that emits `buf_shmem` accesses, then determine how buffer lifetime and overlapping shared storage are represented. Done means generated code uses named base pointers where appropriate while preserving shared-memory reuse and access behavior.

Written by the indexing model from the issue text.