cargo: improve Windows symlink handling

[walter-zeromatter]

Summary

Fixes Windows build failures in the Cargo build script runner caused by incorrect symlink/junction handling. Verified against Zeromatter's repo using --override_module:

• Without this PR (main): build fails with DirectoryNotEmpty (os error 145) when removing junctions, and os error 3 from broken symlink resolution
• With this PR: all symlink-related failures are resolved

Changes

bin.rs

• symlink_if_not_exists now returns Result<bool, ...> — Tracks whether a new symlink was actually created vs one that already existed. Pre-existing entries (e.g. .github junctions that Bazel placed) are no longer added to exec_root_links, so we don't attempt to remove paths we didn't create. This fixes the DirectoryNotEmpty (os error 145) failure observed when crates contain directory entries that become junctions.

• Exec-root link removal is warning-only on Windows — On Windows, symlink removal can fail with PermissionDenied if another process still holds a handle to the target directory (common in parallel Bazel builds). These are temporary links in the build sandbox that Bazel will clean up, so we log a warning and continue rather than failing the entire build script.

• Removed swallow_already_exists helper — The logic is now inlined in the symlink_if_not_exists match arms; the standalone function is unused.

cargo_manifest_dir.rs

• remove_symlink on Windows uses FileTypeExt — The old implementation used path.is_dir(), which follows the symlink target rather than inspecting the reparse point type. The new version uses symlink_metadata() + FileTypeExt to correctly distinguish symlink-files, symlink-dirs, and junctions, then calls the appropriate removal function (remove_file vs remove_dir) with fallbacks.

• Unified drain_runfiles_dir_impl(symlinks_used: bool) — Replaces the separate drain_runfiles_dir_unix and drain_runfiles_dir_windows methods. The key behavioral change: NotFound errors are now tolerated in both modes. On Windows, the runfiles directory can be incomplete (e.g. a Cargo.lock that was never created), and the old code would hard-fail on the missing entry. This was observed as os error 3 failures during builds.

• Added 4 new tests — Cover NotFound tolerance, suffix-retention logic, and the copy-back path for both symlink and non-symlink modes.

Test plan

• Verified main fails building on Windows (DirectoryNotEmpty, os error 3)
• Verified this branch gets past all symlink failures with the same targets
• Linux CI should be unaffected (symlink path is unchanged in behavior, just unified)