loadSupport(): fix global.R support, run global.R in appropriate dir

[alandipert]

• global.R support was broken as a result our changes to the appDir argument; this fixes it.
• I added ye olde require(shiny) so shiny::loadSupport() works.
• Ensures global.R is evaluated in the app directory instead of whichever directory loadSupport() was called in.
• Replace use of on.exit()/setwd() with a withr::with_dir()
• The order of files in R/ is sourced is now given by sort(..., method="radix") which aligns with R conventions
• findEnclosingApp() now uses file.path.ci() internally so it now works on case-insensitive file systems (tested by mounting fat32 loopback disk on Linux)

These items came up preparing https://github.com/CRI-iAtlas/shiny-iatlas for module testing.

[wch]

I think the other files, in the R/ directory, need to be sourced in the app directory as well, though this should be checked first to make sure it's the correct behavior.

[alandipert]

@wch I think you're right, I'll fix it up.

[alandipert]

I did some experimentation with devtools::load_all(), which seems to me like the most analogous existing behavior. I found it to be true that sources in R/ are evaluated with the working dir initially set to the package dir. I also found that the directory is not reset between scripts. That is, one script may call setwd() and influence the working directory of subsequent scripts.

[alandipert]

@wch apprised me of r-lib/roxygen2#1077 and we now sort files from R/ using sort(..., method = "radix") which is documented to use 'C' collation order.