Add instructions on setting a reference level with `formulaic`

[BorisMuzellec]

Is your feature request related to a problem? Please describe.

With the switch to formulaic designs in v0.5.0, the ref_level argument of DeseqDataSet was deprecated. As a result, we may encounter cases in which it is not possible to shrink a given LFC coefficient because it does not correspond to a column in the design matrix. C.f. #361.

Describe the solution you'd like

The reference level can actually be set using the following syntax: design = '~ C(condition, contr.treatment(base="B")).
As this requires some knowledge of formulaic's API, we should

• Add an example in the tutorials
• Update ref_level's deprecation message
• Update lfc_shrink's docstring (and maybe raise a warning when lfc_shrink is called on a non-existant coefficient?)

Describe alternatives you've considered

It might be possible to update the columns of the design matrix without re-fitting dispersions. Something like

dds.design = '~ C(condition, contr.treatment(base="B"))
dds.formulaic_contrasts = FormulaicContrasts(dds.obs, dds.design)
dds.obsm["design_matrix"] = dds.formulaic_contrasts.design_matrix

And then run the pipeline step by step from dds.fit_LFC().

We could consider implementing a method to achieve this.

Additional context
C.f. the discussion in #361.

[ozanozisik]

It will be great also to remind using C() to assure handling of numbers corresponding to categories (e.g. batch) as categorical and not continuous.