relnotes: add 6.5.3

revlog: avoid opening and closing the file for each cloned revision The previous code was flushing files after each new revision, slowing things down. For exemple, with this change, the evolve repository can run `hg debugupgraderepo --run --optimize re-delta-parent` in about 3.4s instead of 4.5 seconds.

censor: accept censored revision during upgrade They can simply be passed by as censored.

censor: show that censored revision prevent repository upgrade This is not great.

smartset: don't ignore hidden revs when intersecting This fixes the bug I demonstrated in the previous commit, but I'm not sure at all if it's the right way of doing it.

tests: demonstrate crash in `unstable()` with internal-phase orphans

rust-matchers: fix quadratic complexity in `FileMatcher` Concretely, this command: ``` $ echo hg up -r <nodeid>; time hg revert dir1 dir2 -r <othernode> --debug hg up -r <nodeid> real 0m14.690s user 0m14.766s sys 0m5.430s ``` was much slower despite using 16 cores before this change. The approach taken here is the same one used in match.py, in exactmatcher. This changeset was originally written by Valentin Gatien-Baron in a private repository. I have redacted the commit message and did a minor clean up of the code.

revlog: add a small cache of unfiltered chunk This can provides a massive boost to the reading of multiple revision and the computation of a valid delta chain. This greatly help operation like `hg log --patch`, delta computation (helping pull/unbundle), linkrev adjustment (helping copy tracing). A first round of benchmark for `hg log --patch --limit 1000` shows improvement in the 10-20% range on "small" repository like pypy or mercurial and large improvements (about 33%) for more complex ones like netbeans and mozilla's. These speeds up are consistent with the improvement to `hg pull` (from a server sending poor deltas) I saw benchmarking this last year. Further benchmark will be run during the freeze. I added some configuration in the experimental space to be able to further test the effect of various tuning for now. This feature should fit well in the "usage/resource profile" configuration that we should land next cycle. When it does not provides a benefit the overhead of the cache seem to be around 2%, a small price for the big improvement. In addition I believe we could shave most of this overhead with a more efficent lru implementation.

revlog: minor refactor in the chunk gather process We will introduce some caching in this method in the next changeset, we make some of the most "disruptive" change first as touching this could break (and maybe did during the development process).

changelog-delay: move the delay/divert logic inside the (inner) revlog Instead of hacking throught the vfs/opener, we implement the delay/divert logic inside the `_InnerRevlog` and `randomaccessfile` object. This will allow to an alternative implementation of the `_InnerRevlog` that does not need to use Python details. As a result, the new implementation can use the transaction less agressively and avoid some extra output since no data had been written yet. That seems like a good side effect.