Implement SIMD funnel shift intrinsics

[Sonike]

Summary

• lower simd_funnel_shl and simd_funnel_shr lane by lane with Cranelift integer operations
• use logical right shifts so signed and unsigned lanes keep the required bitwise semantics
• handle zero shifts explicitly and construct the lane-width constant safely for 128-bit lanes
• re-enable tests/ui/simd/intrinsic/generic-arithmetic-pass.rs in the rustc test suite

The implementation follows the existing scalarized SIMD lowering strategy. Native vector instructions can be considered separately in the future.

Testing

• cargo fmt --check
• ./y.sh build
• ./test.sh
• compiled and ran the upstream generic-arithmetic-pass.rs test against this backend
• compiled and ran focused assertions covering unsigned and signed 32-bit lanes, 128-bit lanes, and shift counts including zero and the maximum valid count

Fixes #1696

[Sonike]

CI note: test_llvm failed in randomized portable-simd f32 simd_min/simd_max property tests involving NaNs and signed zero. This job runs ./y.sh test --use-backend llvm; the workflow explicitly notes that it does not exercise cg_clif. The cg_clif build, targeted upstream regression test, local full suite, and completed cross-platform PR jobs all pass.

I attempted to rerun only the failed job, but GitHub requires repository admin permission for an upstream workflow rerun. A maintainer rerun should confirm whether the randomized LLVM-only test is flaky.