refactor: set do_sample=True if a seed is set on HF backends

[psschwei]

Pull Request

Issue

Fixes #40

Description

HuggingFace backends silently ignored seed (and warned but ignored temperature) because transformers defaults to greedy decoding (do_sample=False). This change forces do_sample=True in _make_backend_specific_and_remove whenever a SEED or non-zero TEMPERATURE is set and the caller hasn't explicitly chosen do_sample, fixing all four HF generate paths in one place.

Testing

• Tests added to the respective file if code was changed
• New code has 100% coverage if code was added
• Ensure existing tests and github automation passes (a maintainer will kick off the github automation when the rest of the PR is populated)

Attribution

• AI coding assistants used

---

Adding a new component, requirement, sampling strategy, or tool?

If your PR adds or modifies one of the types below, check the matching box. A checklist of type-specific review items will be posted as a comment.

• Component
• Requirement
• Sampling Strategy
• Tool

NOTE: Please ensure you have an issue that has been acknowledged by a core contributor and routed you to open a pull request against this repository. Otherwise, please open an issue before continuing with this pull request.

[planetf1]

Minor note on a pre-existing issue this PR slightly expands: _make_backend_specific_and_remove is also called at two apply_chat_template sites (lines 805 and 1087 in huggingface.py), and its output is splatted directly into the tokeniser call. That means temperature was already quietly reaching the Jinja template context before this change — do_sample=True now joins that set.

It does not cause any visible problems today (the templates ignore unknown kwargs), but it is not the right shape — generate-only options should not be in the template namespace. There is already a _filter_chat_template_only_options that strips template-only keys before generate(); a mirror function for the other direction would clean this up. Filed as #1154; nothing needed here.

[psschwei]

gentle ping @generative-computing/mellea-maintainers

[planetf1]

Approved — the /seed coupling is correct and the comment typo is fixed. Deferred cleanup tracked in #1154.