shinyrmd: Safer dependency extraction from pre-rendered HTML

[gadenbuie]

This PR provides a fix for #2499 that doesn't address the root cause but will avoid the downstream errors caused by that issue.

In short, under unknown circumstances, the shinyrmd context script tags for dependencies and execution dependencies are written twice into the pre-rendered HTML files. rmarkdown then passes a length-2 character vector to jsonlite::unserializeJSON(), which causes a parsing error.

This PR does two primary things:

1. I extracted the HTML and JSON parsing of dependencies into a single function shiny_prerendered_extract_context_serialized(). This function ensures a single character string is passed to jsonlite::unserializedJSON(), using the last context in the document if more than one non-unique context is found.

2. rmarkdown reads both dependency types in shiny_prerendered_prerender() (the function that decides if a pre-render run is required). At this point, we don't need to error if parsing fails and can instead trigger a pre-rendered pass to happen again. Ideally, this new pre-render pass will fix any issues in the pre-rendered HTML file, but if it somehow writes invalid dependency JSON again, shiny_prerendered_html() will still fail (rather than infinitely looping).

   Note that avoiding an infinite loop is hypothetical; the current bug wouldn't trigger a new pre-render pass.

[cderv]

This looks good to me! Thanks for this. It makes total sense to do that.

I'd like to know why shiny_prerendered_append_dependencies() is called twice during pre-render

I tried to look into that, but I don't see when we would call render() another time during the same render. shiny_prerendered_append_dependencies() is supposed to append to a file, and this file is supposed to be the knitted file so inermediary results. it would require writing another time to that file and I don't see when we would do that. In general case the intermediary file will be overwritten to a fresh version at each knit() so each render().

So it coulb another render call on the intermediary file but I have check our code base and I don't see that.

I suggest we merge this and this will remain a Mystery