optimize recv_fix_encryption_hierarchy()

recv_fix_encryption_hierarchy() in its present state goes through all
stream filesystems, and for each one traverses the snapshots in order to
find one that exists locally. This happens by calling guid_to_name() for
each snapshot, which iterates through all children of the filesystem.
This results in CPU utilization of 100% for several minutes (for ~1000
filesystems on a Ryzen 4350G) for 1 thread at the end of a raw receive
(-w, regardless whether encrypted or not, dryrun or not).

Fix this by following a different logic: using the top_fs name, call
gather_nvlist() to gather the nvlists for all local filesystems. For
each one filesystem, go through the snapshots to find the corresponding
stream's filesystem (since we know the snapshots guid and can search
with it in stream_avl for the stream's fs). Then go on to fix the
encryption roots and locations as in its present state.

Avoiding guid_to_name() iteratively makes
recv_fix_encryption_hierarchy() significantly faster (from several
minutes to seconds for ~1000 filesystems on a Ryzen 4350G).

Another problem is the following: in case we have promoted a clone of
the filesystem outside the top filesystem specified in zfs send, zfs
receive does not fail but returns an error:
recv_incremental_replication() fails to find its origin and errors out
with needagain=1. This results in recv_fix_hierarchy() not being called
which may render some children of the top fs not mountable since their
encryption root was not updated. To circumvent this make
recv_incremental_replication() silently ignore this error.

Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes #16929