feat: add export_mounts and export_image to auto-export config

[agronskiy]

This tries to address the issue of fast startup of export execution as requested by Gym folks

Approx speedups are here:

Setup	apt-get+git	install	Total
Before (pip, python:3.12-slim)	~1-2 min	~2-3 min	3-5 min
After, cold uv cache	0s	~42s	~42s
After, warm uv cache	0s	~6s	~6s

What was done

• Add export_mounts config key: mount arbitrary host paths (e.g. a pre-heated uv cache) into the export srun container
• Add export_image config key: override the default python:3.12.7-slim export image with a custom one (e.g. one with uv+git pre-installed)

Together these eliminate the ~3-5 min apt-get + pip install overhead in the auto-export step, bringing it down to ~6s with a warm uv cache.

To use this, now use this config (with new launcher installed)

execution:
  auto_export:
    destinations: [mlflow]
    export_image: /path/to/uv-git-python3.12.sqsh  # or a registry image
    export_mounts:
      /lustre/.../users/me/.cache/uv: /cache/uv
    launcher_install_cmd: |
      export UV_CACHE_DIR=/cache/uv
      uv pip install --system "nemo-evaluator-launcher[all] @ git+https://github.com/NVIDIA-NeMo/Evaluator.git@main#subdirectory=packages/nemo-evaluator-launcher"

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[AdamRajfer]

/ok to test fe004b1