gpu: jit: conv_v2: refactor to reduce register usage

- Introduced versioning for offsets to split offset allocations between
  main loop and epilogue
- Added loop_nest_t to simplify work with loop indices/bounds