Raising: buffer normalizations for viewed geps and raw accesses

[wsmoses]

mfem kernels that are not fully inlined reach the raising with buffers addressed through pointer plumbing tensor semantics cannot stand for. This ports the foundation of the normalization suite used to bring the MFEM CUDA suite through the raising path (#2968):

• rebaseViewedGeps: a typed view carved out of a buffer at an offset (gep feeding a pointer2memref) rebases its accesses onto the underlying buffer at base + offset.
• convertRawGepAccesses: data-dependent indexing (CSR-style loops over runtime offsets) can never become affine and stays as raw gep+load; the access still addresses whole elements, so it becomes a plain memref access through a flat view, which raising turns into a gather.
• inlineAllocaScopes / linearizeRegionBlocks: alloca scopes only delimit stack lifetime, which is meaningless under value semantics; splice them into the parent so scratch normalizations see the whole function.
• dropDeadPointerChains: sweep the pointer plumbing the rewrites strand.

Follow-up PRs stack on this: buffer-branch expansion, packed capture-struct forwarding, struct scratch splitting/flattening.

🤖 Generated with Claude Code

https://claude.ai/code/session_016zErYp7upmqr4NHfhod9UD