fix(vmi): lower compact reduction stores with point stores

[mouliangyu]

What

• Normalize full-active continuous stores of compact vcadd/vcmax/vcmin results to unit-stride group_store.
• Treat a full reduction as group = 1 when its only consumer is the store, allowing the existing group-slot lowering to emit 1PT_B32 directly.
• Reuse the same path for explicit grouped reductions with 2, 4, or 8 compact results.
• Preserve ordinary masked-store lowering when the store predicate is dynamic, since point stores cannot represent dynamic predication.

Why

A direct 1 x T -> 1 reduction followed by vstore previously lowered through neutral-init materialization, an extra vector add, predicate construction, and a masked store. The compact result can instead use the existing group-store representation and lower to vcadd plus 1PT_B32 store.

The layout regression test now uses memory-backed operands, allowing layout assignment to select a direct DINTLV_B32 load for multi-part widening/reduction flows instead of introducing an artificial conversion at a VMI register function argument.

Tests

• cmake --build build-main-llvm21 --target pto-test-opt -j2
• llvm-lit -sv build-main-llvm21/test/lit/vmi_new (438/438 passed)

Closes #1022.

The focused reproducer and performance context are in #1013.