fix: ensure augmentation reproducibility with Albumentations RNG

[mandaiy]

Summary

Align custom augmentation randomness with Albumentations-managed RNG so runs can be reproduced via Compose(seed=...).

Changes

• moved ShadesOfGray random sampling into get_params() using self.random_generator
• updated ShadesOfGray.apply() to consume sampled p_norm instead of reading global NumPy RNG
• removed wrapper-level random gate in ultralytics.Albumentations and delegated probability control to Albumentations transforms
• added reproducibility test that validates deterministic outputs when using the same Compose(seed=...)

Testing

• make format
• make lint
• make test
• uv run pytest tests/albumentations/test_transforms.py

Checklist

• I updated relevant documentation under docs/.
• PR title follows Conventional Commits (e.g., feat: ..., fix: ...).
• I verified that the changes are backward compatible, or documented breaking changes.
• I confirmed this PR is ready for review.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4ea5e588f8

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve p argument for Albumentations adapter

Removing the p parameter from Albumentations.__init__ is a backward-incompatible API change: any existing training code that instantiated this wrapper as Albumentations(transform, p=...) will now raise TypeError before augmentation runs. Since this class previously documented and accepted p, keep the argument (even if deprecated and internally forwarded to compose-level probability) to avoid breaking downstream pipelines on upgrade.

Useful? React with 👍 / 👎.