fetch: implement support for atomic reference updates

fetch: implement support for atomic reference updates

When executing a fetch, then git will currently allocate one reference
transaction per reference update and directly commit it. This means that
fetches are non-atomic: even if some of the reference updates fail,
others may still succeed and modify local references.

This is fine in many scenarios, but this strategy has its downsides.

- The view of remote references may be inconsistent and may show a
  bastardized state of the remote repository.

- Batching together updates may improve performance in certain
  scenarios. While the impact probably isn't as pronounced with loose
  references, the upcoming reftable backend may benefit as it needs to
  write less files in case the update is batched.

- The reference-update hook is currently being executed twice per
  updated reference. While this doesn't matter when there is no such
  hook, we have seen severe performance regressions when doing a
  git-fetch(1) with reference-transaction hook when the remote
  repository has hundreds of thousands of references.

Similar to `git push --atomic`, this commit thus introduces atomic
fetches. Instead of allocating one reference transaction per updated
reference, it causes us to only allocate a single transaction and commit
it as soon as all updates were received. If locking of any reference
fails, then we abort the complete transaction and don't update any
reference, which gives us an all-or-nothing fetch.

Note that this may not completely fix the first of above downsides, as
the consistent view also depends on the server-side. If the server
doesn't have a consistent view of its own references during the
reference negotiation phase, then the client would get the same
inconsistent view the server has. This is a separate problem though and,
if it actually exists, can be fixed at a later point.

This commit also changes the way we write FETCH_HEAD in case `--atomic`
is passed. Instead of writing changes as we go, we need to accumulate
all changes first and only commit them at the end when we know that all
reference updates succeeded. Ideally, we'd just do so via a temporary
file so that we don't need to carry all updates in-memory. This isn't
trivially doable though considering the `--append` mode, where we do not
truncate the file but simply append to it. And given that we support
concurrent processes appending to FETCH_HEAD at the same time without
any loss of data, seeding the temporary file with current contents of
FETCH_HEAD initially and then doing a rename wouldn't work either. So
this commit implements the simple strategy of buffering all changes and
appending them to the file on commit.

Signed-off-by: Patrick Steinhardt <ps@pks.im>
Signed-off-by: Junio C Hamano <gitster@pobox.com>