c/partition_allocator: make _highest_group updates monotonic

Previously, in case of failed allocation of partitions for new topics,
we tried to revert _highest_group value as well. When a large number of
topics were allocated in a single request concurrently (e.g. when MirrorMaker
creates required topics on the destination cluster) and some of these
requests bump into partition allocation limits, this could lead to
duplicate group_ids.

As there is no reason to try to reuse group_ids, simply remove the code
moving _highest_group backwards.

Fixes #12610

(cherry picked from commit 3f96297)

src/v/cluster/scheduling/allocation_state.cc

@@ -24,8 +24,6 @@ void allocation_state::rollback(
             remove_allocation(bs, domain);
             remove_final_count(bs, domain);
         }
-        // rollback for each assignment as the groups are distinct
-        _highest_group = raft::group_id(_highest_group() - 1);
     }
 }
 

src/v/cluster/scheduling/partition_allocator.h

@@ -133,7 +133,6 @@ class partition_allocator {
     ss::future<> apply_snapshot(const controller_snapshot&);
 
 private:
-    // reverts not only allocations but group_ids as well
     class intermediate_allocation {
     public:
         intermediate_allocation(

src/v/cluster/tests/partition_allocator_tests.cc

@@ -212,7 +212,7 @@ FIXTURE_TEST(partial_assignment, partition_allocator_fixture) {
 
     BOOST_REQUIRE_EQUAL(3, max_capacity());
     BOOST_REQUIRE_EQUAL(
-      allocator.state().last_group_id()(), max_partitions_in_cluster - 1);
+      allocator.state().last_group_id()(), max_partitions_in_cluster);
 }
 FIXTURE_TEST(max_deallocation, partition_allocator_fixture) {
     register_node(0, 3);