Add constant folding

[nataliakokoromyti]

This PR adds single-pass constant folding for BIR. Folds constant expressions at compile time so 3+4 becomes 7 in the IR instead of an actual add instruction. Handles ints, floats, comparisons, casts, select. One forward pass, DCE cleans up after.

[Zaneham]

I think theres some undefined behaviour here for c99. I think it should be 1ULL in both

[Zaneham]

If a is a negative, left shifting it is undefined behaviour in c99. Needs r = (int64_t)((uint64_t)a << (b & (sw - 1)))

[Zaneham]

I think but happy to discuss other approaches

[Zaneham]

Two more things:

BIR_SDIV / BIR_SREM (the if (b == 0) return 0; guard) -- there's one more edge case here. If sw=64 and a is the most negative 64-bit value and b is -1, the division overflows. Same thing applies to SREM below. Just need a guard to skip folding for that case.

BIR_FPTRUNC (r = (double)(float)r;) -- this always truncates to float32 precision but the target type could be half (f16). Not urgent since constant half expressions are pretty rare, just something to be aware of.

Everything else looks great. No dynamic allocation, deterministic single-pass, correct SSA traversal. Clean work.

Edit: oh, my manners, thank you again!

[nataliakokoromyti]

thank YOU for the feedback and quick response! love BarraCUDA

[Zaneham]

Approved, requested changes have been made, thanks again!