Optimize harfbuzz big integer conversions

Profiling showed that type conversions were adding considerable cycles in time
spent doing text shaping.

The idea is to optimize it using native processor instructions to help Blink
layout performance.

Doing further investigation revelead that compilers may not use the
proper instruction on ARM 32bits builds (i.e. REV16).

One way to insure that the generated ASM was ideal for both gcc/clang
was using __builtin_bswap16.

Added bonus is that we no longer need to test for CPU architecture.

src/hb-machinery.hh

@@ -691,6 +691,10 @@ struct BEInt<Type, 2>
   }
   inline operator Type (void) const
   {
+#if defined(__GNUC__) || defined(__clang__)
+    struct __attribute__((packed)) packed_uint16_t { uint16_t v; };
+    return __builtin_bswap16(((packed_uint16_t *) this)->v);
+#endif
     return (v[0] <<  8)
          + (v[1]      );
   }