A host method's result is read past its width on 32-bit targets

Rows affected, all same after the fix:

• test_nilpy_open_world_method_dispatch: cross True True on i386 (it always was), and mixed True True on arm32 once the softfloat pull changed what the float compare left in r1.
• test_nilpy_pyeval_host_arity_and_returns on arm32: rising(3, 2, 1) came back True.
• test_nilpy_len_of_a_variant_shim_object on arm32: 68719476739 (0x10_0000_0003) for 3, and the truth test read that as True.

Mechanism

TPFn*, TPMI_*, TIFn*, TNoArgI, TVarArgI and TStrArgI are all declared to return Int64, whatever the callee really returns. On x86-64 the callee writes the full register. On 32-bit the second word is whatever the callee last left in r1 or edx.

Fix

PyNarrowRet(v, rk) in pylib.pas narrows a result to its RetKind:

• a byte for Boolean and Char;
• a sign-extended word for Integer;
• on CPU32, the low word for pointer, class and NativeInt kinds.

It is applied at every Int64-thunk result: pyeval's register, mixed and TIFn families, and pylib's PyUserObjNoArgMeth, vi and gi arms.

Test

Rows in test-i386 and test-arm32 for the three tests above. All three are wrong with v446 on arm32 (open-world also on i386).