Docs(gpu): Comprehensive update to README.md

This commit brings the gpu/README.md documentation in line with the
current capabilities and behavior of the install_gpu_driver.sh
initialization action. The previous README was significantly outdated.

Key updates include:

- General:
  - Removed "Beta mode" notice.
  - Updated overall description of the script's purpose and scope.

- Usage and Examples:
  - Added "Default Versions and Supported Configurations" section,
    clarifying default CUDA selection based on Dataproc image version
    and listing supported OS families.
  - Updated `gcloud` command examples to use current best practices
    (regionalized bucket paths, common GPU types like T4).
  - Added new examples for specifying custom driver/CUDA URLs and for
    MIG-enabled cluster creation.
  - Correctly detailed the use of `invocation-type=custom-images`
    metadata, emphasizing its use by image building tools like
    `generate_custom_image.py` rather than direct user specification
    during cluster creation from scratch.

- Metadata Parameters:
  - Completely overhauled and expanded the "Metadata Parameters" section
    to comprehensively list all currently supported options, including:
    - `cuda-version`, `gpu-driver-version`, `cuda-url`, `gpu-driver-url`
      (clarifying HTTP/HTTPS URL input for `cuda-url`/`gpu-driver-url`).
    - `gpu-driver-provider`, `cudnn-version`, `nccl-version`.
    - `install-gpu-agent` (noting new default of `true`).
    - `include-pytorch`, `gpu-conda-env`.
    - `container-runtime`, `http-proxy`.
    - Secure Boot signing parameters (`private_secret_name`, etc.).
  - Provided descriptions, potential values, and defaults for each.

- Feature Documentation:
  - Updated "GPU Scheduling in YARN" to reflect current practices.
  - Added information on NVIDIA Container Toolkit installation.
  - Significantly expanded the "Secure Boot and Kernel Module Signing"
    section with details on MOK management, use of Secret Manager,
    and the `--no-shielded-secure-boot` workaround.
  - Added "Custom Image Creation" section explaining deferred configuration.

- Verification and Notes:
  - Updated verification steps for drivers, CUDA, and the GPU agent.
  - Added an "Important Notes" section covering:
    - OS and Dataproc image version support.
    - Handling of OS package managers and kernel module building.
    - SSHD hardening.
    - Performance considerations, GCS caching mechanism for artifacts,
      the benefits of cache pre-warming (including reduced initial run
      times from up to 150 min to 12-20 min), and the security
      implication of not needing build tools when cache hits occur.
    - Details about the PyTorch Conda environment.

- Removed outdated information regarding very old CUDA versions,
  Dataproc 1.x specific behaviors, and previous GPU agent script handling.

Author: cjac