encoder: fix holdout splitting and compatibility wrapper

[huangyz0918]

Summary

• Preserve callers' benchmark dictionaries when making a new_model holdout split. Previously, removing a target model also removed it from the original input, corrupting later splits made from the same data.
• Repair the backward-compatible prepare_holdout_split() wrapper, which passed an unsupported keyword and raised TypeError. It now uses the equivalent new_pair setting.

These preserve the new-model/new-pair separation described in the paper's experimental setup: https://arxiv.org/html/2604.23099v1#S3.SS1

Verification

• Unchanged repository suite: 123 passed, including PyTorch paths.
• Local checks cover name/index compatibility calls and repeated new-model splits without input mutation.
• Only the production bug fixes are included. Additional validation code stays local; no credentials or live API calls are involved.