fix(infer): align adapter names with batch size

[RerankerGuo]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

Fixes #5285.

Background

TransformersEngine accepted one AdapterRequest for a batch, but
_get_adapter_names() always returned a one-element list. PEFT requires
one adapter name per input, so batches larger than one failed with:

ValueError: Length of `adapter_names` should be the same as the number
of inputs, but got 1 and 10 respectively.

The same mismatch affected base-model inference after an adapter had
already been loaded because the __base__ marker was also returned only
once.

Changes

• Pass the encoded batch size from streaming, forward, and generation
  inference paths to _get_adapter_names().
• Repeat the selected adapter name for every sample in the batch.
• Repeat the __base__ marker when running the base model after loading
  an adapter.
• Add CPU regression tests for both adapter and base-model batches.

Verification

• PYTHONPATH=. .venv/bin/python -m unittest tests.infer.test_transformers_engine_adapters
• PYTHONPATH=. .venv/bin/python tests/run.py --pattern test_transformers_engine_adapters.py
• .venv/bin/pre-commit run --all-files
• git diff --check

Impact

The change is limited to per-batch adapter-name construction in
TransformersEngine. Single-sample inference is unchanged, and the
existing API still applies one AdapterRequest to the whole batch.

Experiment results

Before the fix, both regression cases failed because
_get_adapter_names() could not produce batch-aligned names. After the
fix, the selected adapter and __base__ paths return lists whose lengths
match the batch size; both tests pass.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.