fix(bridge): decode every batch row in generate_stream

[emerardd]

Description

TransformerBridge.generate_stream(..., return_type="str") decoded tokens[0] only, so a batched stream yielded the first sequence's text and the rest of the batch was unreachable. Generation itself was fine — all rows are produced, and return_type="tokens" yields the full [batch, seq_len] tensor — only the decode step dropped them.

This decodes every row and keeps generate()'s unwrap convention: a bare string for a one-row batch, one string per row otherwise. The yield type widens to Generator[Union[torch.Tensor, str, List[str]], None, None].

prompts = ["The capital of France is", "My favourite colour is"]
bridge.generate(prompts, max_new_tokens=4, do_sample=False, verbose=False)
# ['The capital of France is the capital of the', 'My favourite colour is the red. I']

generate_stream(prompts, ...)	Before	After
yield	'The capital of France is the capital of the'	['The capital of France is the capital of the', 'My favourite colour is the red. I']

Single-string and single-row streaming are unchanged, which is why tests/acceptance/model_bridge/test_generate_stream.py::test_stream_returns_strings passed throughout.

HookedTransformer.generate_stream does not share the bug — it never decodes and always yields tensors, even for return_type="str". That is a separate inconsistency, so no mirroring change (AGENTS.md §2) is required here.

Fixes #1756

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

Tests added — both fail on dev and pass with the fix:

• tests/unit/model_bridge/test_generate_stream_batch_decode.py — stubbed token stream, no model load; pins the batched list[str] yield and the single-row bare-string unwrap.
• tests/integration/model_bridge/test_generate_stream_batched.py — distilgpt2 end-to-end; each yield carries one string per prompt and the first yield echoes each row's own prompt.

[jlarson4]

Great work on this @emerardd! Assuming Full Code Coverage passes, I will merge