feat: Add UV support for Ray Client mode

[preneond]

Why are these changes needed?

UV package manager integration currently fails when using Ray Client connections via ray.init("ray://..."). While UV works correctly with the Ray Jobs API (ray job submit -- uv run script.py), it is completely non-functional for interactive Ray Client workflows.

The root cause is that RAY_RUNTIME_ENV_HOOK is explicitly skipped for Ray Client mode (introduced in #26688), which also prevents the built-in UV hook from running. The original skip was necessary because server-side hooks could overwrite GCS URIs with local paths after working_dir upload. However, the UV hook is safe to run client-side before upload.

Fixes #57991

Changes

Core Change: python/ray/util/client/__init__.py

• Added _apply_uv_hook_for_client() function that detects UV environment on the client side before connecting to the cluster
• Modified ClientContext.connect() to apply the UV hook before the server connection
• Custom RAY_RUNTIME_ENV_HOOK remains disabled for Ray Client (safety preserved)

How It Works

1. Client-side: Detects uv run in the local process tree using the existing _get_uv_run_cmdline() function
2. Client-side: Extracts UV args and sets py_executable = "uv run ..." in runtime_env
3. Upload: working_dir (with pyproject.toml/uv.lock) uploads to GCS as normal
4. Cluster: Workers receive py_executable and execute uv run default_worker.py, which installs dependencies from the uploaded lock files

This avoids the issue from #26688 because the hook runs before the working_dir upload (not after), so there's no conflict with GCS URIs.

Tests: python/ray/tests/test_client.py

• test_ray_client_uv_hook_detection — verifies UV detection sets py_executable correctly, handles missing UV, and recovers from errors
• test_ray_client_uv_no_detection_without_uv — verifies normal Ray Client operation is unaffected
• test_ray_client_uv_hook_with_existing_runtime_env — verifies user's working_dir and env_vars are preserved

Tests: python/ray/tests/test_runtime_env_uv_run.py

• test_uv_run_ray_client_mode — end-to-end test verifying UV installs packages via Ray Client

Related issue number

Fixes #57991

Checks

• I've signed off every commit (by using the -s flag, i.e., git commit -s) in this PR.
• [ x] I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
• I've added any new APIs to the API Reference.
• I've made sure the tests are passing (full test_client.py suite: 46/46 passed).
• Testing Strategy
  • Unit tests for _apply_uv_hook_for_client() function
  • Integration tests with Ray Client server
  • Backward compatibility with existing test_env_hook_skipped_for_ray_client

[gemini-code-assist]

Code Review

This pull request adds support for UV package manager in Ray Client mode, which was previously non-functional. The core change involves adding a client-side hook to detect and apply the UV environment before connecting to the cluster. The implementation is well-tested with new unit and integration tests. I've found a couple of issues: an incorrect type hint and a potential bug in how runtime_env is resolved from different sources. My review includes suggestions to fix these.

[gemini-code-assist]

The logic for determining the runtime_env has a potential bug and a misleading comment.

1. The comment Prefer job_config runtime_env if provided is incorrect. The code actually prefers runtime_env from ray_init_kwargs if it's truthy.
2. The expression runtime_env or job_config.runtime_env is problematic when ray_init_kwargs contains an empty dictionary for runtime_env (e.g., ray.init(runtime_env={})). An empty dictionary is a valid value that should take precedence over job_config.runtime_env, but because it's falsy, job_config.runtime_env will be used instead.

To fix this and align with standard Ray precedence rules (ray.init arguments > JobConfig), please make the logic more explicit.

Suggested change

	if job_config and job_config.runtime_env:
	# Prefer job_config runtime_env if provided
	runtime_env = runtime_env or job_config.runtime_env
	if runtime_env is None and job_config and job_config.runtime_env:
	# If runtime_env is not in ray.init(), use the one from JobConfig.
	runtime_env = job_config.runtime_env

[gemini-code-assist]

The return type hint for this function is Dict[str, Any], but it can return None if the input runtime_env is None and the UV hook is not applied. This is confirmed by test_ray_client_uv_hook_detection (Test 3). Please update the type hint to Optional[Dict[str, Any]] for correctness and to aid static analysis.

Suggested change

	def _apply_uv_hook_for_client(runtime_env: Optional[Dict[str, Any]]) -> Dict[str, Any]:
	def _apply_uv_hook_for_client(runtime_env: Optional[Dict[str, Any]]) -> Optional[Dict[str, Any]]: