runtime/internal/atomic: set SP delta correctly for 64-bit atomic functions on ARM

64-bit atomic functions on ARM have the following structure:
- check if the address is 64-bit aligned, if not, prepare a frame
  and call panicUnaligned
- tail call armXXX or goXXX depending on GOARM

The alignment check calls panicUnaligned after preparing a frame,
so the stack can be unwound. The call never returns, so the SP is
not set back. However, the assembler assigns SP delta following
the instruction stream order, not the control flow. So it leaves
a nonzero SP delta after the check, to the tail call instructions,
which is wrong because when they are executed the SP is not
decremented. This CL fixes this by adding the SP back (the
instruction never executes, just tells the assembler to set the
SP delta back).

Should fix #51353.

Change-Id: I976cb1cfb0e9008b13538765cbc7eea0c19c7130
Reviewed-on: https://go-review.googlesource.com/c/go/+/388014


Trust: Cherry Mui <cherryyz@google.com>
Run-TryBot: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>