[kiukchung]

Summary:
A scheduler can now offer more than one workspace builder because MultiWorkspaceMixin selects a builder with workspace_type instead of hard-binding a single mix-in.

class MyScheduler(MultiWorkspaceMixin, Scheduler[Mapping[str, CfgVal]]):
    def __init__(self, session_name: str) -> None:
        super().__init__("my_backend", session_name)
        self._builders = {"dir": DirWorkspaceMixin(), "docker": DockerWorkspaceMixin()}

    def workspace_builders(self) -> Mapping[str, WorkspaceMixin[Any]]:
        return self._builders

torchx run -s my_backend --workspace . \
    -cfg workspace_type=docker,image_repo=example.com/repo utils.echo

slurm now also offers a docker builder that builds and pushes a patched image, renders srun --container-image=<image>, and leaves the img_root macro (the image's root path, substituted into args and env) empty inside the container.

What is unchanged:

• local_cwd still has no workspace support.
• local_docker and kubernetes still use DockerWorkspaceMixin and add no workspace_type option.
• slurm with workspace_type unset still copies into job_dir; the rendered sbatch script stays byte-identical.

Closes #590

Test Plan:

1. pytest across torchx/workspace/test, slurm_scheduler_test, api_test, and workspace_defaults_test -> 130 tests passed.
2. ruff check (import sorting included) and ruff format --check with the pinned ruff -> clean.
3. python -m doctest torchx/workspace/api.py -> passed; pyrefly check -> no findings in the changed files.
4. Downstream plugin suite with torchx installed editable at this commit: 952 passed; the 3 failures are that suite's own generated-file checks and fail identically with the released torchx.