randomize_kstack: Use get_random_u8() at entry for entropy

The randomize kstack code has a couple of sharp edges, which are due to
the fact that a counter was considered more suitable for providing
entropy than the kernel's RNG.

Given that kstack randomization only requires 6-8 bits of entropy,
get_random_u8() is more suitable here, and its fast path has been made
lockless so that there is no overhead beyond a preempt_dis/enable and
a local 64-bit cmpxchg() to update the RNG's internal state.

This means there is no need to defer sampling the counter until syscall
exit, which also removes the need for a per-CPU variable to carry that
value over to the next syscall on the same CPU. That works around some
issues identified by Ryan in [0].

So replace the use of the per-CPU variable in add_random_kstack_offset()
with a direct use of get_random_u8() for all users, and turn
choose_kstack_random_offset() into an empty stub so that each arch can
get rid of it individually.

[0] https://lore.kernel.org/all/dd8c37bc-795f-4c7a-9086-69e584d8ab24@arm.com/

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>