Makefile consolidation and LLVM bug fix

[vtjnash]

The Makefile changes are just to make it easier (more obvious) to handle cases where OpenBLAS doesn't detect the processor type.

The change to alloc.c works around an LLVM optimization bug on Apple's default complier. By re-ordering the layout of the byte cache to be dealing with unsigned indicies, we can safely ignore the sign bit later.

With optimizations on my 64-bit machine, the code originally generated was: (note the movslq instruction which is a 32 to 64 bit sign extend move operation, versus gcc which produces the correct 8 to 64 bit sign extend move instruction movsbq)

with gcc:
0x0000000100046ac0 <+0000> push %rbp
0x0000000100046ac1 <+0001> mov %rsp,%rbp
0x0000000100046ac4 <+0004> movsbq %dil,%rdi
0x0000000100046ac8 <+0008> lea 0x7055d1(%rip),%rax # 0x10074c0a0 <boxed_int8_cache+1024>
0x0000000100046acf <+0015> mov (%rax,%rdi,8),%rax
0x0000000100046ad3 <+0019> leaveq
0x0000000100046ad4 <+0020> retq
0x0000000100046ad5 <+0021> nopl 0x0(%rax,%rax,1)
0x0000000100046ada <+0026> nopw 0x0(%rax,%rax,1)

with llvm:
0x00000001000367c0 <+0000> push %rbp
0x00000001000367c1 <+0001> mov %rsp,%rbp
0x00000001000367c4 <+0004> movslq %edi,%rax
0x00000001000367c7 <+0007> lea 0x6e83f2(%rip),%rcx # 0x10071ebc0 <boxed_int8_cache>
0x00000001000367ce <+0014> mov 0x400(%rcx,%rax,8),%rax
0x00000001000367d6 <+0022> pop %rbp
0x00000001000367d7 <+0023> retq
0x00000001000367d8 <+0024> nopl 0x0(%rax,%rax,1)