mingw: use mimalloc

Thorough benchmarking with repacking a subset of linux.git (the commit
history reachable from 93a6fef ([PATCH] fix the SYSCTL=n compilation,
2007-02-28), to be precise) suggest that this allocator is on par, in
multi-threaded situations maybe even better than nedmalloc:

`git repack -adfq` with mimalloc, 8 threads:

31.166991900 27.576763800 28.712311000 27.373859000 27.163141900

`git repack -adfq` with nedmalloc, 8 threads:

31.915032900 27.149883100 28.244933700 27.240188800 28.580849500

In a different test using GitHub Actions build agents (probably
single-threaded, a core-strength of nedmalloc)):

`git repack -q -d -l -A --unpack-unreachable=2.weeks.ago` with mimalloc:

943.426 978.500 939.709 959.811 954.605

`git repack -q -d -l -A --unpack-unreachable=2.weeks.ago` with nedmalloc:

995.383 952.179 943.253 963.043 980.468

While these measurements were not executed with complete scientific
rigor, as no hardware was set aside specifically for these benchmarks,
it shows that mimalloc and nedmalloc perform almost the same, nedmalloc
with a bit higher variance and also slightly higher average (further
testing suggests that nedmalloc performs worse in multi-threaded
situations than in single-threaded ones).

In short: mimalloc seems to be slightly better suited for our purposes
than nedmalloc.

Seeing that mimalloc is developed actively, while nedmalloc ceased to
see any updates in eight years, let's use mimalloc on Windows instead.

Signed-off-by: Johannes Schindelin <[email protected]>

Author: dscho

config.mak.uname

@@ -750,9 +750,7 @@ ifeq ($(uname_S),MINGW)
 	HAVE_LIBCHARSET_H = YesPlease
 	USE_GETTEXT_SCHEME = fallthrough
 	USE_LIBPCRE = YesPlease
-        ifneq (CLANGARM64,$(MSYSTEM))
-		USE_NED_ALLOCATOR = YesPlease
-        endif
+	USE_MIMALLOC = YesPlease
 	NO_PYTHON =
         ifeq (/mingw64,$(subst 32,64,$(subst clangarm,mingw,$(prefix))))
 		# Move system config into top-level /etc/