Percpu allocations attempt to do first fit by scanning forward from the
first_free of a block. However, fragmentation from allocation requests
can cause holes not seen by block hint update functions. To address
this, create a local version of bitmap_find_next_zero_area_off() that
remembers the largest area skipped over. The caveat is that it only sees
regions skipped over due to not fitting, not regions skipped due to
alignment.

Prior to updating the scan_hint, a scan backwards is done to try and
recover free bits skipped due to alignment. While this can cause
scanning to miss earlier possible free areas, smaller allocations will
eventually fill those holes due to first fit.
Signed-off-by: Dennis Zhou <dennis@kernel.org>