k8s: defer extra pod labels conversion to labels.Selector

[milas]

Extra pod labels come in as one or more maps of label keys and values
([]labels.Set). Once converted into a list of labels.Selector,
it can be used for matching, but introspecting it is more difficult,
as it requires conversion, which produces a list of
labels.Requirement. The biggest problem that introduces is that
labels.Requirement is more expressive, so backwards conversion
means a bunch of error-handling for (what should be) impossible
cases.

This just passes the labels.Sets as they're parsed from the
Tiltfile through and stores them that way on the target in state.
Once they're actually used, they get converted to the actual
labels.Selectors for matching.

This will simplify the new intermediary step where they'll get
converted to API label selector objects, so that conversion can
happen without hiccups or excessive boilerplate.

[milas]

This is just to make the conversion to KubernetesDiscovery easier so that instead of going from labels.Set -> labels.Selector -> metav1.LabelSelector -> labels.Selector, it'll be labels.Set -> metav1.LabelSelector -> labels.Selector.