Break up the percpu counter allocations into individual allocations for
each disk accounting counter; this fixes an issue on large systems where
we have too many replica entries to for the percpu allocator's max
practical size.

Also, use just one eytzinger tree for the normal set of counters and the
gc counters; this simplifies accounting_gc_done() where we need the same
set of counters to be present in both tables.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>