Add Gemma3n model family to fairseq2

[cirquit]

What does this PR do? Please describe:

Adds full Gemma3n (E2B/E4B) support: model architecture, audio tower, tokenizer, and training infrastructure. No vision.

• Decoder with AltUp 4D processing, PLE, LAuReL residuals, QK norm, and KV projection sharing
• USM Conformer audio tower with subsample conv projection and multimodal embedder
• HuggingFace checkpoint import/export and asset card integration
• FSDP (layer/stack), activation checkpointing, and chat template for SFT
• KV projection slot management in decoder (not layer) for FSDP compatibility
• GSM8K SFT config included; verified loss decreases over 200 steps with 8-GPU FSDP

Test:

torchrun --nproc-per-node=8 -m recipes.lm.sft /tmp/gemma3n_sft_test \
      --config-file recipes/lm/sft/configs/gemma3n_e2b_gsm8k.yaml \
      --config trainer.data_parallelism=fsdp \
      --config regime.num_steps=200 \
      --config regime.checkpoint_every_n_steps=100 \
      --config regime.validate_every_n_steps=100

Does your PR introduce any breaking changes? If yes, please list them:
List of all backwards-incompatible changes.

Check list:

• Was the content of this PR discussed and approved via a GitHub issue? (no need for typos or documentation improvements)
• Did you read the contributor guideline?
• Did you make sure that your PR does only one thing instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests?
• Did you verify new and existing tests pass locally with your changes?
• Did you update the CHANGELOG? (no need for typos, documentation, or minor internal changes)

[rsyue]

Really nice to see a clean, proper multimodal implementation via an HG model! Here are some suggestions:

Training config could be a bit more tuned
I do know this is for a quickstart, but wanted to give some input here for a realistic training scenario and for users who might want to start from a baseline.

• fp32 might be a bit overkill. Example: Olmo train config and default Gemma3 causal use bf16 with static mixed precision.
• Default learning rate (3e-4) is quite high for a pretrained model, catastrophic forgetting a possibility.
• LR warmup steps would be useful early in the training process (Olmo config uses 100)
• SDPA default (none) is torch_math, slower and more memory hungry. Consider FlashAttention2 here as a default
• For reproducibility, common.seed would be useful (if training reproducibility is the goal, ignore non-deterministic fa2 comment above)
• Setting max_seq_len or max_num_tokens to correspond to the model defaults or thereabouts would be good in an experimental setting
• Not 100% sure if the checkpoint is still in native hf format, but export_hugging_face should be set to False if so. Gemma3 config disables this (defaults to True)

General questions

• Coordinating chat_mode, currently True. This is a good fairseq2-esque setting, but dataset.py:207 checks for HuggingFaceTokenEncoder and applies the chat template, assuming assistant_masks in output. --> Check if gemma3n tokenizer reliably produces the masks. If I recall, chat_mode is {"src": "Instruction here", "tgt": "Assistant answer here"}
• See comment in the config yaml: I got a KeyError. It seems like dataset.py:181 does a raw dict lookup without error handling. Perhaps chat with team about this, but in the meantime, error handling might be good

Minor but good to have in config

• keep_last_n_checkpoints: 10 with ckpt every 100 steps could miss out on better previous ckpts, if this is used in a baseline setting

[rsyue]

LGTM!

[zyaoj]

LGTM! Some nit comments.

[cirquit]

Thanks to everybody for contributing!