Require explicit opt-in for SpeechLM2 remote code

[pzelasko]

What does this PR do?

Fixes #15599 by making trust_remote_code an explicit runtime decision for SpeechLM2 HuggingFace Hub loads.

• Overrides any checkpoint-configured trust_remote_code value with the caller's runtime argument (default: False).
• Removes trust_remote_code from exported checkpoint configurations so the decision is not persisted.
• Replaces the remaining hardcoded trust_remote_code=True in DuplexEARTTS._load_language_model.
• Documents explicit opt-in with an immutable reviewed revision.
• Adds regression coverage for malicious config opt-in, explicit caller opt-in/opt-out, non-persistence, and the DuplexEARTTS language-model path.

Why?

A checkpoint controls its own config.json, including the tokenizer and language-model repository names. Allowing that same configuration to enable remote code execution means an untrusted model can opt itself into importing its repository Python code during load. The trust decision must come from the loading application instead.

Verification

• pre-commit run --from-ref origin/main --to-ref HEAD
• git diff --check origin/main...HEAD
• Python syntax compilation for the changed Python files
• Focused isolated regression harness covering runtime opt-in/opt-out and checkpoint non-persistence

The full NeMo pytest dependency environment was not available locally; repository CI will run the dependency-backed tests.

[pzelasko]

/ok to test d40d07f

[Edresson]

LGTM

[github-actions]

[🤖]: Hi @pzelasko 👋,

We wanted to let you know that a CICD pipeline for this PR just finished successfully.

So it might be time to merge this PR or get some approvals.