In the generic atomic ops that rely on a loop around a CAS primitive,
there's no need to force the initial read of the "old" value to be atomic.
In the typically-rare case that we get a torn value, that simply means
that the first CAS attempt will fail; but it will update "old" to the
atomically-read value, so the next attempt has a chance of succeeding.
It was already being done that way in pg_atomic_exchange_u64_impl(),
but let's duplicate the approach in the rest.

(Given the current coding of the pg_atomic_read functions, this change
is a no-op anyway on popular platforms; it only makes a difference where
pg_atomic_read_u64_impl() is implemented as a CAS.)

In passing, also remove unnecessary take-a-pointer-and-dereference-it
coding in the pg_atomic_read functions.  That seems to have been based
on a misunderstanding of what the C standard requires.  What actually
matters is that the pointer be declared as pointing to volatile, which
it is.

I don't believe this will change the assembly code at all on x86
platforms (even ignoring the likelihood that these implementations
get overridden by others); but it may help on less-mainstream CPUs.

Discussion: https://postgr.es/m/13707.1504718238@sss.pgh.pa.us