Raising: forward the packed capture struct; distribute geps over selects

[wsmoses]

Stacks on #3007 (which stacks on #3006).

A not-quite-inlined device lambda packs its captures (DeviceTensors: pointer + sizes) into a stack struct and reads them back through typed views at constant offsets. Forward each load to the unique dominating store of its slot so the pointer-typed members never reach the raising as memory:

• constant-attribute and constant-valued-SSA gep indexing;
• wide integer stores covering several slots (two i32 sizes packed into one i64 write, extracted little-endian shift+trunc);
• runtime-indexed shape reads (sizes[i] walked in a loop, as a memref.load or an affine map like s0 floordiv 4) resolved as a select chain over the constant-offset stores of their element type;
• stranded dead pointer selects (left by the buffer-branch expansion) skipped — narrowly, since a general dead-user skip admits big register scratch whose forwarding is unsound.

Also distributeGepOverSelect: a side-selected view (select of byte offsets or of the pointers themselves feeding geps) distributes the gep and the memref view over the select so the buffer-branch expansion (#3007) can take over.

This is what makes mfem's bilininteg_vecdiffusion_pa.cpp and the batched LOR kernels raise fully (#2968).

🤖 Generated with Claude Code

https://claude.ai/code/session_016zErYp7upmqr4NHfhod9UD

[wsmoses]

Recreated targeting main (content rebased): the stack lock prevents changing this PR's base.