[alexreinking]

This PR fixes two bugs in prefetch lowering. Both showed up when prefetching the weights in the GGML app (#9511).

Struct-typed buffers. StorageFlattening emitted a struct-typed prefetch call. Struct-typed Exprs may only appear as field()'s argument or as a Store's value, so compilation failed. The prefetch is now of the buffer's bytes: the offset and strides are in bytes, and the record's bytes are added as the innermost dimension.

Lets that depend on the loop. InjectPrefetch computes the prefetched box by advancing the loop variable by the offset in the body under the prefetch. Lets between the loop and the prefetch kept their values from the current iteration. Bounds inference creates such lets, for example for a realization computed inside a GuardWithIf tile. In the test, a tile of 4 rows reads 6 rows of the input, but the prefetch covered only 2 rows (and none in the tail). Those lets are now re-wrapped around the body, so they are evaluated at the prefetched iteration too.

correctness_prefetch gains two tests, run on all of its targets:

• test13 prefetches a struct-typed buffer. Without this PR it fails to compile.
• test14 checks the GuardWithIf case above. Without this PR it finds 2 prefetched rows instead of 6.

Validation:

• ctest -R 'correctness_(prefetch|struct_type)$' on this branch: passed.
• ctest -L correctness at the top of the stack (apps/ggml: GGML quantized kernels on Approximations #9511): 704/705 passed. The failure is correctness_cross_compilation, which fails because my LLVM build has no Hexagon target.

Stacked on #9599.

Checklist

• Tests added or updated (not required for docs, CI config, or typo fixes)
• Documentation updated (if public API changed)
• Python bindings updated (if public API changed)
• Benchmarks are included here if the change is intended to affect performance.
• Commits include AI attribution where applicable (see Code of Conduct)