In particular, this enables HTTP/2 support in reqwest, which is a lot
more efficient.

The FunctionCallTrace object consumes a few hundred bytes of stack
space, even when tracing is disabled. This was causing stack overflows:

  $ nix-instantiate '<nixpkgs> -A texlive.combined.scheme-full --dry-run
  error: stack overflow (possible infinite recursion)

This is with the default stack size of 8 MiB.

Putting the object on the heap reduces stack usage to < 5 MiB.

This is no longer needed since we're not using POSIX locks anymore.

(cherry picked from commit 3392f1b7)

(cherry picked from commit 96c6b08e)

(cherry picked from commit 63c5c91c)

(cherry picked from commit 8beedd44)

(cherry picked from commit fb692e5f)

This allows writing attribute lookups as

    if (auto name = value.attrs->get(state.sName))
      ...

(cherry picked from commit f216c76c)

It's now regenerated when util.hh changes, and is ordered after
config.h to fix a race.

(cherry picked from commit d0a769cb)

(cherry picked from commit a0de58f4)

(cherry picked from commit f70434b1)

(cherry picked from commit 15a16e5c)

This reverts commit 2b761d5f.

Also *really* make fmt() take arguments by reference.

doc/manual: add note to `allowSubstitutes` advanced attribute

https://hydra.nixos.org/build/107724274

This makes it possible to reference single attribute definitions,
for pointing people to their exact definition.

Fix typos

Make nix/unpack-channel.nix a builtin builder

This is a hack to fix the build on macOS, which was failing because
libnixrust.a contains compiler builtins that clash with
libclang_rt.osx.a. There's probably a better solution...

https://hydra.nixos.org/build/107473280

Closes #3225.
Closes #3226.

Fixes #3246.

https://hydra.nixos.org/build/107467517

Seems that on i686-linux, gcc and rustc disagree on how to return
1-word structs: gcc has the caller pass a pointer to the result, while
rustc has the callee return the result in a register. Work around this
by using a bare pointer.

https://hydra.nixos.org/build/107466992

Also, fetchGit now runs in O(1) memory since we pipe the output of
'git archive' directly into unpackTarball() (rather than first reading
it all into memory).

In particular, these are emitted by 'git archive' (in fetchGit).