model: add RBLNModernBertForMaskedLM

[ho4040]

Type of Change

• New Model Support

Changes Overview

Adds RBLNModernBertForMaskedLM (+ RBLNModernBertForMaskedLMConfig) for the
ModernBERT encoder, following the existing RBLNRobertaForMaskedLM pattern.

ModernBERT needs two model-specific adjustments over the generic encoder path:

• Force SDPA at load (get_pytorch_model). ModernBERT selects its attention
  backend from config._attn_implementation and prefers HuggingFace's
  flash_attention_2 (the CUDA-only flash-attn kernels) when that package is
  present; rebel-compiler cannot compile that path for RBLN. We pin
  attn_implementation="sdpa" so compilation is deterministic regardless of the
  host environment and the compiler lowers scaled_dot_product_attention to its
  own kernel. (This is unrelated to RBLN's own flash_attn attn_impl, which
  is a decoder KV-cache feature and does not apply to encoders.)
• A dedicated compile wrapper (ModernBertModelWrapper) forwards the 2D
  padding mask unchanged instead of pre-expanding it with
  _prepare_4d_attention_mask. ModernBERT alternates full and local
  (sliding-window) attention and builds both 4D masks internally via
  create_bidirectional_mask / create_bidirectional_sliding_window_mask;
  pre-expanding would feed the global mask to the sliding layers and drop the
  local windowing.

The class is registered in the three __init__ export lists;
RBLNAutoModelForMaskedLM resolves it through the existing naming convention.
A test mirroring TestBertForMaskedLM is added
(hf-internal-testing/tiny-random-ModernBertForMaskedLM).

Motivation and Context

ModernBERT (including the multilingual mmBERT) currently has no official
optimum-rbln support. We built and validated this adapter end-to-end on real
ATOM+ hardware:

• Env: ATOM+ (RBLN-CA22), rebel-compiler 0.11.0, optimum-rbln 0.11.0.post1,
  transformers 5.8.1, torch 2.11.0
• Compile: jhu-clsp/mmBERT-base, export=True, max_seq_len=128,
  batch_size=1, fp32 → compiled_model.rbln produced
• Accuracy vs CPU eager (SDPA): top-1 agreement = 1.00 over all real
  token positions; masked-token prediction matches ("Paris"); masked-slot
  logit cosine similarity 0.957

We also published a technical write-up of the port, showcasing running an
unsupported encoder on ATOM+:

• English: https://warpspaceinc.github.io/blog/posts/modernbert-on-rebellions-npu/
• 한국어: https://warpspaceinc.github.io/blog/ko/posts/modernbert-on-rebellions-npu/

We'd love ModernBERT to become officially supported so we can point readers at
first-party support. Happy to adjust to your conventions, extend sequence-length
/ bucketing coverage, or hand off for an accuracy review.

Related Issues

[rebel-kblee]

LGTM 👍 Thanks for contribute!