io_uring/msg_ring: add an alloc cache for CQE entries

io_uring/msg_ring: add an alloc cache for CQE entries

io_uring accounts the memory allocated, which is quite expensive. Wrap
the allocation and frees in the provided alloc cache framework. The
target ctx needs to be locked anyway for posting the overflow entry,
so just move the overflow alloc inside that section. Flushing the
entries has it locked as well, so io_cache_alloc_free() can be used.

In a simple test, most of the overhead of DEFER_TASKRUN message passing
ends up being accounting for allocation and free, and with this change
it's completely gone.

Signed-off-by: Jens Axboe <axboe@kernel.dk>