[Tracking] Align rsync-based sync with Git semantics

[andylizf]

We should make rsync-based syncing (workdir, file_mounts) match Git behavior as closely as practical, i.e. consistent ignores, ownership, symlinks across all sync methods.

Context

workdir and local file_mounts still rely on rsync, which many developer workflows depend on. But rsync behavior differs from Git in several key areas (ignores, ownership, symlinks), causing inconsistencies.

• Git workdir support ([Core] Support git repo #6257) gives us a clean baseline where remote state matches git clone
• We have .skyignore ([Storage] Add .skyignore support #4038), but with issues reported in [core] inconsistent .gitignore behavior with sky launch rsync #5492
• Earlier cleanup attempt in Consolidate file filtering logic between rsync and bucket upload #5006 (closed as stale)

Problems

1. .skyignore lacks negation pattern support (!pattern)

The current --filter='dir-merge,- .skyignore' implementation strips Git-style negation patterns. We should either port to the storage_utils parser for full Git-style support, or at least detect and forbid unsupported syntax with clear error messages.

2. Nested .skyignore files ignored without root file

We only enable .skyignore mode if the source root has one; otherwise everything falls back to .gitignore. This contradicts docs that say "Any .skyignore … is respected." Should either honor nested files or warn users clearly.

3. Root/Docker ownership inconsistencies

When the remote login user is root, rsync -a preserves local UID/GID on remote files. This differs from Git checkout and from non-root clusters. Should force --no-owner --no-group under root, or at least document/warn.

---

Note: Still need to check if COPY and storage_mounts have similar issues (#5006).

[andylizf]

Cc @cblmemo @Michaelvll for a look here

[Michaelvll]

Thanks for raising this @andylizf! It is a known issue, and we need to figure out a efficient way to support git alike ignore behavior. : )