topK/bottomK edge behavior mismatch between thanos engine and the native engine

[alanprot]

The topK and bottomK operators return errors on the native engine when the first argument (k) is empty on all cases and in the Thanos engine, we only return an error if the second argument is not empty.

Ex: the following expression returns error on the native engine but not in the new engine:

bottomk(scalar(do_not_exist), do_not_exist)`

Adding a simple check to return an error always if the arg is nil does not work as it breaks some distributed queries tests where some shards can be empty.

[MichaHoffmann]

I wonder: is topk/bottomk even distributable? bottomk(X, Y) if X resides on one engine and Y resides on another seems broken, right? We probably only should distribute if the k is a *parser.NumberLiteral? In that case we can just add the check, right?

[MichaHoffmann]

Does this testcase make sense?

		{
			name:     "topk non-distributable",
			expr:     `topk(scalar(X), Y)`,
			expected: `topk(scalar(X), dedup(remote(Y), remote(Y)))`,
		},
		{
			name:     "topk distributable",
			expr:     `topk(1, Y)`,
			expected: `topk(1, dedup(remote(topk by (region) (1, Y)), remote(topk by (region) (1, Y))))`,
		},