[AIEX] Implement foldImmediate Peephole optimization

[khallouh]

Fold move-immediate + COPY pairs into a single move-immediate at the
consumer site:
%c = MOVA 42 → %a = MOVA 42
%a = COPY %c
Guarded by hasOneNonDBGUse to avoid materializing constants multiple
times when they have several consumers, which would inflate register
pressure and perturb allocation.

[khallouh]

Top improvements Topk1D_int8_0, LayerNormC8Part2_aie2_int8_0 and SinBf16_0 (7%, 5% and 3% respectively in Instruction Count) There are a few regressions but most are small, highest being 1.5% in BufferPadInnermost. These are mostly due to side effects on decisions made by the register allocator, so not really easily recoverable.

Edit: Topk1D_int8_0 number is not reliable because its runtime takes random paths in each run, which brings down our top improvement to 5% in LayerNormC8Part2_aie2_int8_0