Tell armv7l assembler this is a constant

Because armv7 opcodes are only 32 bits, `mov` can only move immediates
that are 8 bits or less.  There's some wiggle room with shifting and
whatnot, but in full generality you need to actually perform a load from
a memory address.  The assembler will insert the immediate into a
literal pool, but that pool has to be <4KB away from the current `pc` at
the time of load, so we add `.ltorg` to flush the literal pool at the
end of every function.

Author: staticfloat

src/trampolines/trampolines_arm.S

@@ -6,14 +6,15 @@
 .cfi_startproc; \
 MANGLE(UNDERSCORE(name))##:; \
     ldr ip, CONCAT(.L,NAMEADDR(name)); \
-    mov r12, SYMBOL_IDX(idx); \
+    ldr r12,=#SYMBOL_IDX(idx); \
 CONCAT(.L,MANGLE(UNDERSCORE(name))): ;\
     add ip, pc, ip; \
     ldr pc, [ip]; \
     .align 2; \
 CONCAT(.L,NAMEADDR(name))##: ; \
     .word NAMEADDR(name)##-(CONCAT(.L,MANGLE(UNDERSCORE(name))) + 8); \
 .cfi_endproc; \
+.ltorg; \
 
 // Generate both `dgemm_` and `dgemm_64_`
 #include "ilp64_doubling.h"