[vince62s]

Bump setup.py and eole.version to 0.6.7, and align the README/changelog, Docker image references, RunPod workflow guidance, repository template, release-check fixtures and maintainer checklist.

This patch release includes merged #501: startup selects a compatible DFlash2 checkpoint/precision before preload and warmup, matching the workspace default and avoiding the first default request unloading warmed MTP and compiling DFlash2 again. Missing/incompatible drafters retain native settings; later explicit mode/precision switches remain supported.

Validation:

• Full CPU suite: 812 tests, OK (43 skipped).
• Full docs/build_docs.sh passed: Sphinx with warnings as errors, TypeScript checking and the production Docusaurus build including anchor validation.
• Black, flake8 and whitespace checks passed.
• Clean tracked-source builds produced eole-0.6.7.tar.gz and eole-0.6.7-py3-none-any.whl, both passing Twine checks.
• Verified package version agreement, all workspace wheel assets, DFlash2 license, CUDA sources in the source distribution and RunPod target.

After merge:

1. Tag the merged release commit 0.6.7 and publish its GitHub release for PyPI/docs.
2. Build Docker Images with source_ref=0.6.7, Torch 2.14.1 and CUDA 13.2.1.
3. Build RunPod Image with source_ref=0.6.7, deploy_tag=0.6.7 and ghcr.io/eole-nlp/eole:0.6.7-torch2.14.1-ubuntu24.04-cuda13.2.1 (or its digest).
4. Validate the new images on a GPU, including DFlash2 startup/first request, then update the existing public template e9e19ozt9c to ghcr.io/eole-nlp/eole-runpod:0.6.7.

The startup fix is full-image server code: rebuild the parent as well as the RunPod wrapper. No tag/release was published, image workflow triggered or live template changed by this PR.