feat: add NemotronH hybrid model support with multi-GPU VRAM calibration

NemotronH (Mamba2 SSM + MoE + Attention) requires several changes to
load and abliterate correctly on multi-GPU systems.

Architecture support (model.py):
- Add backbone.layers fallback in get_layers() for NemotronH's
  model.backbone.layers structure
- Add get_layer_modules() patterns for NemotronH's unified mixer
  attribute: mixer.out_proj (Mamba2), mixer.o_proj (attention),
  mixer.down_proj / mixer.experts[*].down_proj /
  mixer.shared_experts.down_proj (MoE)
- Scan all layers in get_abliterable_components() instead of only
  layer 0, to discover the full union of component types in hybrid
  architectures
- Add _get_hidden_states_via_hooks() fallback for models that don't
  return hidden_states through generate() (NemotronH returns tuple
  of Nones); use forward hooks on each layer with device-aware
  stacking for multi-GPU compatibility
- Skip meta-device and NaN-weight modules in abliterate() to prevent
  NaN corruption when layers are CPU-offloaded by Accelerate
- Add _has_mamba_layers() to detect hybrid SSM architectures

Multi-GPU VRAM calibration (model.py):
- After inference warmup on multi-GPU systems, check if any GPU has
  less than 6 GiB free; if so, release the model, measure actual free
  VRAM per GPU, and reload once with corrected per-GPU caps
- Overloaded GPUs get a 0.7 correction factor for Accelerate's
  layer-size underestimation; other GPUs get full budget to absorb
  displaced layers; gated to hybrid SSM models via _has_mamba_layers()
  so regular transformers are unaffected

User experience:
- Show trust_remote_code explanation with model repo link before
  prompting, replacing the bare HuggingFace error message
- Auto-install mamba-ssm when required, with clear nvcc/CUDA toolkit
  guidance on build failure
- Suggest installing causal-conv1d and mamba-ssm after loading any
  model with Mamba layers when fast kernels are missing

Other fixes:
- Sum VRAM across all GPUs in print_memory_usage() (utils.py)
- Show total and per-GPU VRAM in startup output (main.py)
- Fix division by zero in evaluator when base_refusals is 0
- Add mamba optional dependency group to pyproject.toml

Author: cpagac

src/heretic/evaluator.py

@@ -110,7 +110,9 @@ def get_score(self) -> tuple[tuple[float, float], float, int]:
         kl_divergence_scale = self.settings.kl_divergence_scale
         kl_divergence_target = self.settings.kl_divergence_target
 
-        refusals_score = refusals / self.base_refusals
+        refusals_score = (
+            refusals / self.base_refusals if self.base_refusals > 0 else 0.0
+        )
 
         if kl_divergence >= kl_divergence_target:
             kld_score = kl_divergence / kl_divergence_scale

src/heretic/main.py

@@ -54,14 +54,22 @@
 )
 
 
-def obtain_merge_strategy(settings: Settings) -> str | None:
+def obtain_merge_strategy(settings: Settings, model: Model) -> str | None:
     """
     Prompts the user for how to proceed with saving the model.
     Provides info to the user if the model is quantized on memory use.
     Returns "merge", "adapter", or None (if cancelled/invalid).
     """
 
-    if settings.quantization == QuantizationMethod.BNB_4BIT:
+    # Also detect pre-quantized models (FP8, MXFP4, etc.) via their built-in
+    # quantization_config, which HuggingFace stores in the model's config.json.
+    pre_quantized = (
+        getattr(model.model.config, "quantization_config", None) is not None
+        and settings.quantization == QuantizationMethod.NONE
+    )
+    is_quantized = settings.quantization == QuantizationMethod.BNB_4BIT or pre_quantized
+
+    if is_quantized:
         print()
         print(
             "Model was loaded with quantization. Merging requires reloading the base model."
@@ -753,7 +761,7 @@ def count_completed_trials() -> int:
                             if not save_directory:
                                 continue
 
-                            strategy = obtain_merge_strategy(settings)
+                            strategy = obtain_merge_strategy(settings, model)
                             if strategy is None:
                                 continue
 
@@ -802,7 +810,7 @@ def count_completed_trials() -> int:
                             )
                             private = visibility == "Private"
 
-                            strategy = obtain_merge_strategy(settings)
+                            strategy = obtain_merge_strategy(settings, model)
                             if strategy is None:
                                 continue