x86: Improve the trampoline in 64-bit mode

At present this leaves the stack at the pre-relocation value. This is not
ideal since we want to have U-Boot running entirely from the top of
memory.

In addition, the new global_data pointer is not actually used, since
the global_data pointer itself is relocated, then the pre-relocation value
is changed, so the effective value (after relocation) does not update.

Adjust the implementation to follow the 32-bit code more closely, with a
trampoline function which is passed the new stack and global_data pointer.
This ensures that the correct values come through even when relocating.

Change-Id: Id9b0fb3198dfc1b7aa6ed0380f4b8f77edc473d6
Signed-off-by: Simon Glass <sjg@chromium.org>