Add stop_strings and stopping_criteria support to TransformerBridge.generate

[RecreationalMath]

Description

• Adds stop_strings and stopping_criteria to TransformerBridge.generate() and generate_stream(), alongside the existing stop_at_eos. HuggingFace-native stopping is already reachable via hf_generate(). This brings the same stopping surface to the bridge's own generation loop, which previously supported only EOS stopping.
• stop_strings reuses HuggingFace's StopStringCriteria rather than a hand-rolled matcher, so its matching behavior is identical to HuggingFace.
• The three stop signals (EOS via stop_at_eos, stop_strings, and stopping_criteria) are independent: any one of them ends a sequence.
• Supported on the standard decoder-only text path (single and batched). Encoder-decoder, inputs_embeds, and multimodal inputs always raise a clear NotImplementedError. Stateful/SSM models raise only when run with use_past_kv_cache=False (their default keeps them on the supported path).

Notes for review

• Backward compatibility: with both new parameters left as None (the default), generation output is byte-for-byte identical to before, so existing callers are unaffected.
• The stop_at_eos=False interaction required a change to shared loop logic, so the diff touches the existing EOS-handling path: the early-exit and finished-row padding were previously gated on stop_at_eos, and are now gated on whether any stop signal is active. Without this, a stop_strings match with stop_at_eos=False would mark a sequence finished but never actually end generation.
• One test exercises the default use_past_kv_cache=True path and is skipped on macOS-arm64 (the upstream KV-cache NaN, issue [Bug Report] [macOS-arm64] Cached eager attention NaNs in transformers v5 — blocks bridge KV-cache generation #1322), registered in tests/QUARANTINES.md per repo policy. Every other test runs on all platforms.

Tests added

A new integration suite covering single and list stop_strings, custom criteria (bare, list, and StoppingCriteriaList), the stop_at_eos=False case, no-op defaults, batched generation, generate_stream parity, and the error contracts.

Closes #595

Type of change

• New feature (non-breaking change which adds functionality)
• This change requires a documentation update

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

[RecreationalMath]

Both red checks are the known flakes, not this PR: Notebook Checks (BERT) (import-cell stderr mismatch, passes locally) and Docstring Test (HF-429 on the model-loading doctests).

[jlarson4]

Excellent work on this! False negative CI tests have passed on a re-run, merging as is