[mono][jit] Fix float to int32 casting overflow behavior

[jandupej]

The issue has been encountered in #85163

The code private static float Z(float y) => (float)(int)y; translates on arm64 to:

...
000000000000001c        fcvtzs  x0, s0
0000000000000020        sxtw    x0, w0
0000000000000024        scvtf   s0, w0
...

Note the x0 in fcvtzs. The float value is cast to int64, then the lower half gets sign-extended. The expected behavior would be to directly convert to int32 with saturation, i.e.:

...
000000000000001c        fcvtzs  w0, s0
0000000000000020        scvtf   s0, w0
...

This would be more in line with the casting behavior of CoreCLR. This would also fix the test /JIT/SIMD/VectorConvert_ro_Target_64Bit/, which is to be reenabled at that point.

[jandupej]

Fixed in the PR above.