Fixes #112871

In lowering, we had logic that inserted an intermediate cast to int when casting small types to/from floating. The float->small case was already handled by morph and will not appear in lowering. The other direction doesn't require a cast as long as the small type is handled correctly in the containment check and codegen.

The extra cast that was being inserted broke small->uint->floating conversions, for which uint->floating is now handled directly by AVX-512 instructions.

Diffs show a few hits where the uint->floating conversion instruction is restored, but all cases are on loads of unsigned small types where it wouldn't matter.

        movzx    rax, byte  ptr [rbp-0x0F]
        vxorps   xmm0, xmm0, xmm0
-       vcvtsi2sd xmm0, xmm0, eax
+       vcvtusi2sd xmm0, xmm0, eax

I've added a regression test for the Fuzzlyn example, which does fail without the fix.

There are a few improvements as well. I removed a restriction on containment of memory operands for ulong->floating casts when we can use the AVX-512 conversion instructions:

-       mov      rax, qword ptr [rdi]
        vxorps   xmm0, xmm0, xmm0
-       vcvtusi2ss xmm0, xmm0, rax
+       vcvtusi2ss xmm0, xmm0, qword ptr [rdi]

And we're now skipping the insertion of unnecessary int casts, which eliminates some movs:

-       mov      eax, esi
        vxorps   xmm1, xmm1, xmm1
-       vcvtsi2ss xmm1, xmm1, eax
+       vcvtsi2ss xmm1, xmm1, esi