[gemma4] feat: add Gemma-4 31B dense model support

[Zhichenzzz]

Stacked on #7 (Gemma-4 MoE HF-faithfulness fixes).

Ports upstream NVIDIA-NeMo/Megatron-Bridge#3885 ("limited Gemma4 dense support") onto this fork, adapted to our fork's norm fusion, adding Gemma-4 31B-it (dense, non-MoE) RL support.

Changes

gemma4_provider.py

• Add attention_k_eq_v: bool = False field.
• _install_tied_kv: gate on attention_k_eq_v (covers both MoE and dense) instead of num_moe_experts is None. Both 26B-A4B-it and 31B-it set attention_k_eq_v=True, so this is a no-op for the 26B MoE path.

gemma4_vl_bridge.py

• Unblock the dense path for models without per-layer embeddings (hidden_size_per_layer_input == 0, e.g. 31B-it). Dense models with PLE still error since MCore lacks PLE support.
• Branch provider config on enable_moe_block: dense path sets num_moe_experts=None and ffn_hidden_size=intermediate_size. MoE path unchanged.
• Pass provider.attention_k_eq_v = text_config.attention_k_eq_v.
• Add dense weight mappings (inert on MoE — those keys don't exist there):
  • mlp.linear_fc1.weight via GatedMLPMapping(gate_proj, up_proj)
  • mlp.linear_fc2.weight ← down_proj
  • mlp.linear_fc1.layer_norm_weight ← pre_feedforward_layernorm.weight — differs from upstream [model] feat: Add limited Gemma4 dense model support NVIDIA-NeMo/Megatron-Bridge#3885, which maps it to post_attention_layernorm. Our fork fuses post_attention_layernorm into linear_proj.post_layernorm (TERowParallelLinearLayerNorm), so the MLP's fused fc1 norm corresponds to HF's MLP input norm (pre_feedforward_layernorm), not post_attention_layernorm. Wrong norm → high lpdiff; lpdiff ~0.007 confirms this mapping.

Validation

31B-it dense (e2e RL on dapo-math-17k, 8× H200, TP=4, no expert parallelism, n=8 / response 768 for reward variance):

• 0 missing / 0 unexpected keys on load.
• Coherent rollouts (real math reasoning).
• train_rollout_logprob_abs_diff ≈ 0.007 (range 0.004–0.008, well under the 0.02 target).
• Stable through real gradient updates: weight_version 1 → 2 → 3, lpdiff stays in the 0.004–0.008 band (no blowup).

26B-A4B-it MoE regression check on this branch: lpdiff 0.0086 — matches the validated baseline from #7. No regression.

Notes

• Launch scripts for 31B dense ship in the miles PR (radixark/miles#1232).
• Stacked on Gemma-4: HF-faithful MoE fixes (dual pre-norm, global rotary, final-logit softcapping) #7. When Gemma-4: HF-faithful MoE fixes (dual pre-norm, global rotary, final-logit softcapping) #7 merges to bridge, this PR's base will retarget to bridge and the visible diff narrows to the dense delta.