If a shader's only use of float16 is for FConvert, add the Float16 capability.

When visiting instructions, check that there are no 16-bit storage capabilities
and add the Float16 capability. Same for int8/int16.

Test/spv.float16convertonlyarith.comp

@@ -0,0 +1,11 @@
+#version 450 core
+
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+
+layout(local_size_x = 16, local_size_y = 16, local_size_z = 1) in;
+
+void main()
+{
+    vec4 v = vec4(0.0);
+    vec4(f16vec4(v));
+}