variations on treatment name in `lmitt`

This is related to the new wiki page on [Functionality we aim to support](https://github.com/benbhansen-stats/flexida/wiki/Functionality-we-aim-to-support).

In this issue I'm only referring to use of the `lmitt()` function, not the `as.lmitt(lm(...))` path.

1) Should we force there to be a variable on the RHS of the equation? In other words, while `lmitt(y ~ assigned(), ...)` works, should we allow either `lmitt(y ~ 0, ...)` or `lmitt(y ~ 1, ...)` as a short-hand? (Or some other version? I wish `lmitt(y ~ , ...)` worked but it unfortunately errors.)

2) Should we allow the use of the treatment variable by name, rather than `assigned()`? I'd vote no on this, if only because it makes things much more complicated.

3) If we disallow using the treatment variable name in lieu of `assigned()`, what should something like `lmitt(y ~ z, ...)` produce? An error? A warning and replace `z` with `assigned()`? Currently it creates the `z:assigned()` interaction.

4) What about just not allowing any identification of treatment on RHS? Subgrouping variables only.

(There's some scattered discussion of these ideas in other issues; bringing in a clear issue *only* for the topic of identifying treatment on RHS of `lmitt()`.)

Start with the lmitt() entry point and read the linked wiki page, then review the scattered discussion in related issues about identifying treatment on the right-hand side. Compare the four proposed behaviors and determine which treatment-expression rules should be adopted. Done means the behavior is decided clearly enough to guide an implementation.

Written by the indexing model from the issue text.