pyoxidizer: produce working Python 3 Windows installers (issue6366) While we've had code to produce Python 3 Windows installers with PyOxidizer, we haven't been advertising them on the web site due to a bug in making TLS connections and issues around resource handling. This commit upgrades our PyOxidizer install and configuration to use a recent Git commit of PyOxidizer. This new version of PyOxidizer contains a *ton* of changes, improvements, and bug fixes. Notably, Windows shared distributions now mostly "just work" and the TLS bug and random problems with Python extension modules in the standard library go away. And Python has been upgraded from 3.7 to 3.8.6. The price we pay for this upgrade is a ton of backwards incompatible changes to Starlark. I applied this commit (the overall series actually) on stable to produce Windows installers for Mercurial 5.5.2, which I published shortly before submitting this commit for review. In order to get the stable branch working, I decided to take a less aggressive approach to Python resource management. Previously, we were attempting to load all Python modules from memory and were performing some hacks to copy Mercurial's non-module resources into additional directories in Starlark. This commit implements a resource callback function in Starlark (a new feature since PyOxidizer 0.7) to dynamically assign standard library resources to in-memory loading and all other resources to filesystem loading. This means that Mercurial's files and all the other packages we ship in the Windows installers (e.g. certifi and pygments) are loaded from the filesystem instead of from memory. This avoids issues due to lack of __file__ and enables us to ship a working Python 3 installer on Windows. The end state of the install layout after this patch is not ideal for @: we still copy resource files like templates and help text to directories next to the hg.exe executable. There is code in @ to use importlib.resources to load these files and we could likely remove these copies once this lands on @. But for now, the install layout mimics what we've shipped for seemingly forever and is backwards compatible. It allows us to achieve the milestone of working Python 3 Windows installers and gets us a giant step closer to deleting Python 2. Differential Revision: https://phab.mercurial-scm.org/D9148

tests: set git config using `git config` for simplicity I also took the liberty to not set the same config value twice as we did before. Differential Revision: https://phab.mercurial-scm.org/D9176

tests: use `git init` instead of unusual synonym `git init-db` I had never heard of `git init-db` until I saw our tests. Differential Revision: https://phab.mercurial-scm.org/D9175

errors: name arguments to Abort constructor Differential Revision: https://phab.mercurial-scm.org/D9179

errors: stop passing non-strings to Abort's constructor The next patch will change `Abort`'s constructor and `__bytes__` functions and they will start assuming that the first argument is the messages as `bytes`. Differential Revision: https://phab.mercurial-scm.org/D9178

errors: name arguments to UnknownCommand constructor Differential Revision: https://phab.mercurial-scm.org/D9166

errors: name arguments to AmbiguousCommand constructor Differential Revision: https://phab.mercurial-scm.org/D9165

errors: name arguments to CommandError constructor Because readability counts. Differential Revision: https://phab.mercurial-scm.org/D9164

errors: move UnknownCommand and AmbiguousCommand near CommandError They seem closely related. Differential Revision: https://phab.mercurial-scm.org/D9163

templatekw: make {successorssets} always return a list (issue6342) Previously, {successorssets} returns an empty string instead of an empty list for a non-obsolete changeset. The changing type of the JSON output makes it hard to consume from statically-typed languages. Differential Revision: https://phab.mercurial-scm.org/D9158