compiler: saturate float-to-int conversions

This works around some UB in LLVM, where an out-of-bounds conversion would produce a poison value.
The selected behavior is saturating, except that NaN is mapped to the minimum value.

testdata/float.go

@@ -29,7 +29,12 @@ func main() {
 	var f2 float32 = 5.7
 	var f3 float32 = -2.3
 	var f4 float32 = -11.8
-	println(int32(f1), int32(f2), int32(f3), int32(f4))
+	var f5 float32 = -1
+	var f6 float32 = 256
+	var f7 float32 = -129
+	var f8 float32 = 0
+	f8 /= 0
+	println(int32(f1), int32(f2), int32(f3), int32(f4), uint8(f5), uint8(f6), int8(f7), int8(f6), uint8(f8), int8(f8))
 
 	// int -> float
 	var i1 int32 = 53

testdata/float.txt

@@ -11,7 +11,7 @@
 +3.333333e-001
 +6.666667e-001
 +6.666667e-001
-3 5 -2 -11
+3 5 -2 -11 0 255 -128 127 0 -128
 +5.300000e+001 -8.000000e+000 +2.000000e+001
 (+6.666667e-001+1.200000e+000i)
 +6.666667e-001