Fix optimized armhf builds

Apparently gcc-6.2 with -O3 and -mfpu=neon can use ARM instructions
that requires 64bit alignment (here vst1.64) with data that
is not 64bit aligned (g->strcache) https://i.imgur.com/vYktsUn.png

So we need to be able to build "arm_neon_functions.cc" with
-mfpu=neon, while not automatically using NEON for the rest
of the codebase, unless explicitly asked for.

wscript

@@ -86,6 +86,8 @@ compiler_flags_dictionaries= {
         'attasm': '-masm=att',
         # Flags to make AVX instructions/intrinsics available
         'avx': '-mavx',
+        # Flags to make ARM/NEON instructions/intrinsics available
+        'neon': '-mfpu=neon',
         # Flags to generate position independent code, when needed to build a shared object
         'pic': '-fPIC',
         # Flags required to compile C code with anonymous unions (only part of C11)
@@ -120,6 +122,7 @@ compiler_flags_dictionaries= {
         'c99': '/TP',
         'attasm': '',
         'avx': '',
+        'neon': '',
         'pic': '',
         'c-anonymous-union': '',
     },