fix: Windows cross-compilation — unsigned long vs size_t in polynomial_arithmetic.hpp (#22072)

## Summary

Fixes the second Windows cross-compilation failure on v4-next (after PR
#22056 fixed the first one in file_io.hpp).

The header `polynomial_arithmetic.hpp` declares `unsigned long
num_coeffs` but the implementation (.cpp) uses `const size_t
num_coeffs`. On Linux these are the same type, but on Windows LLP64
`unsigned long` is 32-bit while `size_t` is 64-bit, causing
`-Wshorten-64-to-32` when passing `domain.size` (size_t).

This function was removed on `next` before Windows cross-compilation was
added, so it was never caught there. It's v4-specific legacy code.

CI failure:
https://github.com/AztecProtocol/aztec-packages/actions/runs/23608075427/job/68756199559

ClaudeBox log: https://claudebox.work/s/fe9af88896e7e6d6?run=8

Author: ludamad

barretenberg/cpp/src/barretenberg/polynomials/polynomial_arithmetic.hpp

@@ -87,7 +87,7 @@ void coset_ifft(std::vector<Fr*> coeffs, const EvaluationDomain<Fr>& domain);
 // src_domain, const EvaluationDomain<Fr>& target_domain);
 
 fr compute_barycentric_evaluation(const fr* coeffs,
-                                  unsigned long num_coeffs,
+                                  size_t num_coeffs,
                                   const fr& z,
                                   const EvaluationDomain<fr>& domain);