Verify data in mesh header is consistent.

It seems that in some files the triangles are considered part of the
header, and included in the header length. The skins then follow the
header, pretty much always. In fact header_len can be treated as skin_start,
which does seem like a more likely interpretation. It may be worth making the
same assumption with the main file header, where header_len is in fact
frame_start. This makes the whole thing look more sane.

Extract normals:

<damien> heres python for decoding normal from a 16bit value
<damien> lat = (latlng >> 8) & 0xFF;
<damien> lng = (latlng) & 0xFF;
<damien> lat *= math.pi/128;
<damien> lng *= math.pi/128;
<damien> x = math.cos(lat) * math.sin(lng)
<damien> y = math.sin(lat) * math.sin(lng)
<damien> z =                 math.cos(lng)
<damien> retval = [ x, y, z ]
<damien> thats how quake 3 does it apparently

Implement freeing models.

Check for an use sinf and cosf if available.

Freeing mesh in libmd3_mesh_load is bad, because it wasn't allocated
as a mesh.