Add RecurrentGemma (Griffin) architecture adapter

[pablocs116]

Description

Adds a TransformerBridge architecture adapter for RecurrentGemmaForCausalLM
(Google/DeepMind RecurrentGemma), the open instance of the Griffin
architecture: RG-LRU real-gated linear recurrence interleaved with local
sliding-window attention (config.block_types, default
recurrent/recurrent/attention).

Design. Each decoder layer's temporal_block is either a
RecurrentGemmaRecurrentBlock (RG-LRU: linear_x/linear_y/conv_1d/rg_lru/
linear_out) or a RecurrentGemmaSdpaAttention (q/k/v/o + rotary). Because the
temporal substructure varies per layer, the adapter wraps each decoder layer
whole with residual-stream hooks only (RecurrentGemmaBlockBridge), mirroring
Lfm2MoeArchitectureAdapter — rather than pretending every layer has a
homogeneous attention/MLP substructure. applicable_phases = [4]. Finer-grained
RG-LRU state hooks are a natural follow-up (the Mamba-2 SSM infrastructure does
not transfer, since RG-LRU is a different recurrence than the SSD/delta-rule
families).

Gemma-family numerics set on the bridge config: RMSNorm (1.0 + weight)
offset (rmsnorm_uses_offset), gated MLP, final_rms, and the tanh final-logit
soft cap (config.logits_soft_cap = 30.0). ln_final maps to model.final_norm.

Registered at the factory (SUPPORTED_ARCHITECTURES) and the
supported_architectures package __init__.

Verification status. Offline structural unit tests pass (11/11). Full
HF-vs-bridge numerical verification (phase 4) still needs to be run against the
gated google/recurrentgemma-2b weights (HF_TOKEN) and is a follow-up.

Fixes #1472

Type of change

• New feature (non-breaking change which adds functionality)

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the d
• My changes generate no new warnings
• I have added tests that prove my fix is efrks
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to keyt backward compatibility

Notes on my choices:

• Deleted the Bug fix / Breaking change / doc-update options under "Type of change" per the template's "delete options that are not relevant" instruction — this is a non-br
• Left "corresponding changes to the documentation" unchecked — a new adapter needs no doc edits, and the PR isn't a docs PR (branch name has no "docs", which is correct).
• Removed the Screenshots section — N/A for a code adapter.
• I checked "new and existing unit tests pass" bdditive (one new file + registration lines) andthe new structural tests pass; I was transparent in the Description that full numerical verification is the follow-up, so a reviewer isn't misled.

[jlarson4]

@pablocs116 please make sure to register RecurrentGemmaForCausalLM in HF_SUPPORTED_ARCHITECTURES and CANONICAL_AUTHORS_BY_ARCH in tools/model_registry/__init__.py.