MLX DPO, ORPO and SimPO cut the prompt to max_length // 2 even when the row fits (follow-up to #1209) #2199

Context

Blocked by #1209 and #1196. Please do not start before both have landed.

The maintainer ruled on #1209 that DPO and ORPO on the transformers backend cut the prompt only when
prompt plus answer overflow max_length; #1196 puts the same rule on SimPO, KTO, IPO and BCO. The
review threads of both left the MLX files out on purpose, as a separate change.

The MLX trainers mirror TRL's tokenization so that one soup.yaml trains on the same tokens on
either backend (see the module docstring of src/soup_cli/trainer/mlx_dpo.py). Today they cut the
prompt by the prompt alone:

• DPO: src/soup_cli/trainer/mlx_dpo.py:160 (_truncate_tokens(prompt_ids, max_prompt_length, ...)
  with max_prompt_length=max_length // 2 from _build_example, :504-507).
• ORPO: src/soup_cli/trainer/mlx_orpo.py:162 (build_orpo_example, called with
  max_length // 2 from _build_example, :313-316). mlx_simpo.py builds its rows with the ORPO
  wrapper, so SimPO shares this site.

Once #1209 and #1196 are in, a long-prompt row that fits trains on different tokens on MLX and on
transformers.

Proposed fix

Apply the same rule in the two MLX builders: cut the prompt only when the prompt plus the answer
exceed max_length, from the same end as today (DPO keeps the start of the prompt, ORPO and SimPO
the end), and let the answers keep the room that remains. Reuse the rule #1196 puts in
trainer/_trl_compat.py if it works on token lists, otherwise a small pure function with the same
cases. The row-building tests run without Apple hardware (fake tokenizer); a live MLX run is a
welcome extra for someone with Apple Silicon, not a requirement.

Acceptance

• A DPO row and an ORPO/SimPO row that fit in data.max_length keep their whole prompt in build_preference_example and build_orpo_example; a test fails if the max_length // 2 cap touches them.
• A row that overflows is still cut from the same end as today.
• The preference-prompt wording in docs/backends-and-ops.md says both backends cut under the same rule.
• A changed changelog fragment says long-prompt MLX configs now keep the whole prompt when the row fits.

Found during the v0.77.0 release review.