Fix the test target not compiling against Quick 7 and libgit2 1.x

[goneng]

Why this is needed

objective-git ships 207 unit tests. None of them have run in years,
because the target they live in does not compile. CI never says so: the
Test project step has been commented out in BuildPR.yml:50 since
"Split test and archive in CI" (Feb 2025), so only Archive project
runs and a dead test target looks exactly like a healthy one.

That is not a theoretical gap. #107 fixed a stack overrun in
GTRepository+Merging.m, where a git_merge_file_input was initialized
with git_merge_file_options_init, memcpy'ing a 48-byte template over a
40-byte struct. GTRepositorySpec.m:264 calls the very method that
contained it. A test target that built would have had a real chance of
catching it, and the same class of bug can walk in again tomorrow.

This PR does not enable tests in CI and does not make the suite pass. It
makes the target build and the suite run, which is the prerequisite for
both, and it is a prerequisite for #80 as well.

What was broken

Three independent things, each hiding the next:

1. No valid architecture. The vendored xcconfig sets
   VALID_ARCHS = x86_64 for Mac targets. The framework target
   overrides it with "x86_64 arm64"; the test target never did. On
   Apple Silicon the target produced no binary at all and the build
   died at Build input file cannot be found, before compiling a
   single source file.

2. git_buf.asize is gone. libgit2 1.x moved the allocated size
   into reserved. The three NSDataGitSpec assertions keep their
   original meaning, because git_buf_grow still writes the allocation
   size into that field.

3. The specs were never migrated to Quick 7. QuickSpecBegin now
   expands to + (void)spec, so self in a spec body is the spec
   class, not the running example, while the fixture helpers are
   instance methods on QuickSpec. Fourteen of these were hard errors.
   The other 79 compiled only because clang permits any known selector
   on a bare Class, and would have raised unrecognized selector sent to class at run time, so fixing only the 14 would have produced a
   target that builds and then dies on the first example.

What still blocks a green run

Worth knowing before anyone turns the CI step back on. Neither is caused
by this PR, and both are better handled separately:

• fixtures.zip cannot be extracted by a modern ZipArchive. It holds 66
  absolute symlinks under repo-with-submodule/.git/modules/** that
  point into /Applications/GitHub.app/..., left by the GitHub for Mac
  client that produced the fixture in 2013. ZipArchive 2.6.0 rejects any
  symlink escaping the destination and aborts the whole extraction, so
  no fixture repository is ever created. Worse, the half-extracted
  directory is then treated as a valid cache by
  setUpRepositoryFixtureIfNeeded:, so the unzip is never retried.
• Nimble's beginWith force-casts the actual value to String
  (BeginWith.swift:49), so expect(nil).to(beginWith(...)) aborts the
  whole xctest process instead of failing one example.

Test plan

• xcodebuild -workspace ObjectiveGitFramework.xcworkspace -scheme "ObjectiveGit Mac" build-for-testing ARCHS=arm64 gives
  ** TEST BUILD SUCCEEDED **, exit 0, zero errors. On master the same
  command fails with Build input file cannot be found.
• test-without-building then runs the suite for the first time: 207
  tests executed, where master executes none. They fail, every one of
  them traceable to the fixture extraction described above, which is
  the next thing to fix.
• Reverting only the Quick commit and keeping the other two reproduces
  the 14 incomplete definition of type 'struct objc_class' errors, so
  each commit is doing work that is actually needed.