Determine how to handle water tracers with negative qflx_qrgwl, and generally how to handle water tracers for fixed-volume lakes

As I understand it, qflx_qrgwl is used to balance the water budget in landunits where this isn't inherently balanced. For example, in lakes, a positive `P - E` leads to positive qflx_qrgwl, whereas a negative `P - E` leads to a negative qflx_qrgwl. A negative `P - E` can be sustained over time due to biases is the modeled P or E, or because we don't model river influxes into lakes.

I think we need to do some thinking around how to handle water tracers when we have a negative qflx_qrgwl. Consider lakes, for example. If `P - E` is positive, I think we can simply use the tracer concentration from precipitation to set the tracer concentration in the positive qflx_qrgwl. But if `P - E` is negative, it's unclear how we should set the tracer concentration in the negative qflx_qrgwl: Tracer concentration depends on the source of the water, and at the point when the negative qflx_qrgwl is generated, we don't know what the source of this will be (I think possibilities include some mix of offsetting other positive fluxes, river water from ROF, ocean water if the river is unable to accommodate this negative flux, or the global redistribution of the negative fluxes that's done in CMEPS to avoid negative runoff fluxes). This is problematic for setting the tracer flux associated with qflx_qrgwl, and also (I think) for setting the tracer concentration of lakes and other water bodies that are "refilled" using qflx_qrgwl.

My initial thinking about how to resolve this follows [my thinking for irrigation](https://github.com/ESCOMP/CTSM/issues/512#issuecomment-4898296928): We should never have negative fluxes like this. Instead, we should have explicit pools of water in CTSM that are refilled as much as possible through positive fluxes from ROF that are sent in response to a demand term. For lakes, we already have such pools, but a difference would be that these would become variable in volume, which means that they could in principle go to zero; in this case, I guess evaporation would be shut down until the lake could be refilled. (I think we could possibly keep this volume of water separate from the virtual water volume that's used for the calculation of lake energy fluxes: we could possibly keep depths constant for those purposes.) The mapping of the demand term from LND to ROF and its subsequent satisfaction could be done in any way (it would not need to conserve exactly because it is not itself a flux). In particular, the demand term could be handled in a way that it is almost guaranteed to be satisfied (at the cost of some complexity) by doing an ever-expanding search in ROF for rivers that can satisfy this demand.

I'm not sure if this initial thinking makes any sense. In any case, this warrants more thought.