Add FORBIDDEN_TOKENS to Gemma4Tokenizer covering image + audio placeholders

[dparikh79]

Summary

Gemma4Tokenizer did not define FORBIDDEN_TOKENS, so it inherited the base class default of (). The sampler at gemma/gm/text/_sampler.py:501 does:

forbidden_tokens = self._normalize_tokens(self.forbidden_tokens)
forbidden_tokens += self.tokenizer.FORBIDDEN_TOKENS

For Gemma 4 this added nothing, leaving the sampler free to emit raw multimodal placeholder tokens (<|image|>, <start_of_image>, <image|>, <|audio|>, <start_of_audio>, <audio|>) during text-only inference. The result is corrupted text-only output that includes raw placeholder ids the user never asked for.

Fix

Mirror the existing Gemma3Tokenizer / Gemma3nTokenizer pattern on the new class, but include all six distinct multimodal ids. Gemma 3 reuses IMAGE_PLACEHOLDER == START_OF_IMAGE == 255999, so listing both there is redundant; in Gemma 4 all six are distinct per _Gemma4SpecialTokens (verified directly in this file):

IMAGE_PLACEHOLDER = 258880  # <|image|>
START_OF_IMAGE = 255999     # <|image>
END_OF_IMAGE = 258882       # <image|>
AUDIO_PLACEHOLDER = 258881  # <|audio|>
START_OF_AUDIO = 256000     # <|audio> (BOA)
END_OF_AUDIO = 258883       # <audio|> (EOA)

so all six need to be forbidden to keep the sampler from emitting any of them in text-only mode.

Test Plan

• Added test_gemma4_tokenizer_forbids_multimodal_placeholder_tokens in gemma/gm/text/_tokenizer_test.py. It iterates over all six expected multimodal ids and asserts each is present in Gemma4Tokenizer.FORBIDDEN_TOKENS. This is a class-attribute test, so it does not need to load the actual SentencePiece model.
• Verified via AST that the new FORBIDDEN_TOKENS tuple contains exactly the six expected special_tokens.* entries.
• No regression for Gemma3Tokenizer / Gemma3nTokenizer (their FORBIDDEN_TOKENS is unchanged).
• No em-dashes / no AI-tells / Google CLA already signed (covered via the mujoco / dm_control umbrella).

Fixes #613.

Credit to @ac12644 for the diagnosis and proposed fix in the issue.