Raising: expand buffer-yielding branches into their accesses

[wsmoses]

Stacks on #3006.

mfem's coefficient ternary (const_coeff ? c(0,0,0) : c(i,j,e)) arrives at the raising as a branch yielding a pointer or a whole memref — as an arith.select of buffers, an scf.if yielding a pointer whose arms compute geps, or an affine.if yielding a memref. No tensor value can stand for the yielded buffer, so each access is pushed down into a clone of the branch: only scalars cross the yield and the branch raises as a select. Arms may contain loads (reads are idempotent and clone safely per pushed-down access); branches whose buffer results die are rebuilt without them when scalar yields keep the branch alive.

The existing raise_buffer_select golden changes from a whole-tensor select to a per-access select — equivalent, and the expansion also covers the cases the whole-buffer path had to reject (a store through the selected buffer).

Part of #2968.

🤖 Generated with Claude Code

https://claude.ai/code/session_016zErYp7upmqr4NHfhod9UD

[giordano]

Stacks on #3006.

Quoting myself from JuliaMath/SpecialFunctions.jl#552 (comment)

Claude loves to talk about stacking PRs, but not actually stacking them

Edit: switched to a real stack.