merge: use in-memory mergestate when using in-memory context This is my version of Augie's D8568. It makes it so we don't touch the mergestate on disk when using an in-memory context. The reason that I want this is not the same as the reason that Augie write his patch (though I agree with that reason too). My hope is to make in-memory rebase not fall back to on-disk rebase when there are conflict. I plan to do that by adding a `overlayworkingctx.reflect_in_workingcopy()`. The idea is that that will update the working copy, the dirstate and the mergestate as necessary. Differential Revision: https://phab.mercurial-scm.org/D9040

mergestate: extract a base class to be shared by future memmergestate This extracts a new base class from `mergestate` and leaves all the vfs-touching code in `mergestate`. Differential Revision: https://phab.mercurial-scm.org/D9039

mergestate: extract overridable methods for making/restoring file backups In-memory merge currently still uses on-disk mergestate. That's mostly harmless (I don't think I've seen any problem it's actually caused). It's still a little weird. I'm planning to add an in-memory mergestate, which will be used with `overlayworkingctx`. This patch prepares for that by extracting that logic, so it's easier to change per subclass. Differential Revision: https://phab.mercurial-scm.org/D9038

mergestate: initialize all properties in __init__() This is hopefully not very controverial. I found the initialization before this patch unorthodox. It wasn't clear which properties the object was supposed to have. Differential Revision: https://phab.mercurial-scm.org/D9037

mergestate: remove unnecessary clearing of `localctx` and `otherctx` As noted in the previous commit, there are no callers that (re-)use the instance after calling `reset()`. There are also no callers that call `_read()` after doing anything with the instance (it's only called right after an instance is created). If reviewers feel that this is too risky, I can extract and reuse the poisoning code that indygreg once added for poisining repo instances. Differential Revision: https://phab.mercurial-scm.org/D9036

mergestate: move most of of reset() into start() `ms.reset()` has somehow become a combination of two different things: 1. A constructor-like function creating an empty instance 2. A call to `shutil.rmtree()` to clear the mergestate. It seems that all callers now care only about the latter (since we changed one caller to use the new `ms.start()` method instead). Let's move the code for the former into `start()`, since that's the only place it's needed. Differential Revision: https://phab.mercurial-scm.org/D9035

mergestate: make clean() only be about creating a clean mergestate This is similar to the previous patch, but moves the separation one step further out, to `merge.py`. Differential Revision: https://phab.mercurial-scm.org/D9034

mergestate: split up reset() for its two use cases We only have one place that calls `ms.reset()` with any arguments. That place is `mergestate.clean()`. The callers that call the function with no arguments seem to all just want delete the mergestate -- none of them look at the instance after calling `reset()`. Let's separate out the two different use cases to make the code clearer. I'll clean up further soon. Differential Revision: https://phab.mercurial-scm.org/D9033

mergestate: simplify reset(), knowing that `other` and `node` go together There's only one caller of `reset()` that passes any arguments at all, and that originates from `merge.py:1371`. That code always passes values for both `node` and `other`. Differential Revision: https://phab.mercurial-scm.org/D9032

mergestate: make some callers not pass pointless node argument The node argument is set on the created `mergestate` instance, but these callers don't even look at that instance. Differential Revision: https://phab.mercurial-scm.org/D9031