Problem:  The test suite takes over 17 minutes serially and cannot
          be run with make -j because tests share the testdir/
          working directory and create temporary files with
          well-known names (X*, test.log, messages, vimcmd).
Solution: Create per-test isolated working directories under
          workdir/ using symlink trees that mirror the real
          directory layout, keeping temp files isolated while
          preserving all relative path assumptions.
          (Jesse Rosenstock)

Each test runs in workdir/$test/testdir/ where the symlink tree
preserves relative-path depth: ../ resolves to src/ contents and
../../ resolves to the vim root.  The layout is:

  workdir/                     symlinks to vim root (runtime/, ...)
  workdir/$test/               symlinks to src/ contents
  workdir/$test/testdir/       symlinks to testdir/ contents, test cwd

Level-1 symlinks are created once in the nolog target rather than
per-test, because ln -s follows existing symlinks to directories
(creating e.g. runtime/runtime in the repo root) and there is no
race-free way to create them from parallel test processes.  A
setup_workdir.sh POSIX shell script creates the per-test levels
2 and 3.

The Makefile avoids GNU Make extensions to stay POSIX-compatible:
setup_workdir.sh replaces a define/call macro, shell TESTDIR=`pwd`
replaces $(CURDIR), and recursive $(MAKE) replaces order-only
prerequisites (|).

Tests that navigate directory structure (test_stacktrace.vim,
test_python3.vim) are adapted for the extra workdir/ level.

An alternative approach would eliminate the symlink tree entirely
by rewriting all ~127 relative path references (../ and ../../)
across 35 test files to use absolute paths via g:testdir.  This
was rejected because: (1) it touches many files unrelated to the
parallelism fix, creating review noise and merge conflicts;
(2) tab-completion and fnamemodify tests genuinely test relative
path resolution against real directory structure, requiring
per-test fixture directories anyway; (3) future test authors
must remember to avoid ../ in favor of g:testdir, an easy
regression with no enforcement mechanism; and (4) the symlink
approach preserves the natural relative-path idiom that existing
tests use, keeping the diff focused on infrastructure.

On a 48-core machine, wall clock time drops from 17m32s to ~50s.

This supersedes 6146f33 / vim#20082 (patch 9.2.0410: test suite races when
run with parallel make) which added .NOTPARALLEL as a workaround for the
same race conditions.

related: vim#20082

Signed-off-by: Jesse Rosenstock <jmr@google.com>
Co-authored-by: Gemini <noreply@google.com>