Allow objects with a _repr_html_() method to be rendered to HTML

[cpsievert]

As we've pointed out in several py-shiny/py-shinywidget issues (e.g., posit-dev/py-shiny#303), rendering "ad hoc (i.e., non-ipywidget) widgets" like folium or itables is possible, but it isn't easy since to do because you need ui.HTML(x._repr_html_()) to get the relevant HTML.

This PR makes it so objects that implement a _repr_html_ method are renderable when they appear as a child of a Tag/TagList. Also note that, using this approach, we can statically and dynamically render these objects:

import folium

from shiny import App, render, ui

app_ui = ui.page_fixed(
    "Map 1",
    folium.Map(location=(45.5236, -122.6750)),
    ui.br(),
    "Map 2",
    ui.output_ui("map")
)

def server(input, output, session):
    @output
    @render.ui
    def map():
        return folium.Map(location=(45.5236, -122.6750))

app = App(app_ui, server)

[wch]

This is added inside of the else code block, but I think it would be clearer if it were added at the same level as the else. Something like this makes it really clear what's going on (and the reader doesn't have to look at _normalize_text to find out that it's a no-op for HTML objects):

            elif isinstance(child, ReprHtml):
                if prev_was_add_ws:
                    html_ += "  " * indent

                html_ += HTML(child._repr_html_())  # type: ignore[reportPrivateUsage]

                prev_was_add_ws = False

One more thing I wonder about: the code as currently written assumes that the _repr_html_ method does not want to have space around it. But in practice, I think that objects with _repr_html_ are always block elements, because they're meant to be displayed beneath a notebook cell. So perhaps the whitespace logic should be changed to reflect that? Or are there any cases of _repr_html_ where they want to be inline?

[cpsievert]

Done in fb78cda. I agree that most if not all _repr_html_ cases should be block elements, but I don't think it's terrible to say x._repr_html() gets the same treatment as HTML(), especially considering that adding whitespace correctly would add considerable complexity here without obvious gain (we can't control whitespace within x._repr_html(), so producing readable HTML is mostly up to x).

[wch]

I just mean that if we treat it as block element, we'd do this:

            elif isinstance(child, ReprHtml):
                html_ += "  " * indent

                html_ += HTML(child._repr_html_())  # type: ignore[reportPrivateUsage]

                prev_was_add_ws = True

[cpsievert]

I think what you're wanting would be more involved to implement correctly, though. With your suggestion, there is nothing to handle to eol whitespace (here's a minimal example):

import htmltools as h

class Foo:
    def _repr_html_(self) -> str:
        return "<span>Foo</span>"

print(str(h.span(f)))

<span>  <span>Foo</span></span>

As I said before, it doesn't feel worth the added complexity for such a minor detail. Plus, I like that the behavior is consistent with HTML()

[wch]

The behavior in your example (using the code I suggested) is the same as the current behavior if you put a block (_add_ws=True) element inside of an inline (_add_ws=False) element.

import htmltools as h
print(str(h.span(h.div())))

<span>  <div></div></span>

Do you think we should expect some different behavior here?

[cpsievert]

Oh yea, I do find that surprising, but it doesn't seem like a big problem since that's not valid HTML