Add IBM Granite architecture support

[Bhavyashah20]

Adds architecture definition for GraniteForCausalLM (IBM Granite 3.x dense models). Without this, mergekit logs "No JSON architecture found for GraniteForCausalLM" and falls back to inference for every Granite model.

Granite uses the same transformer weight layout as Llama (q/k/v/o projections, gate/up/down MLP, input and post-attention layernorms) with model_type "granite". Optional bias entries are included for models released with attention_bias=True or mlp_bias=True to prevent silent tensor loss during merges.

Adds make_picogranite() to test helpers and TestGraniteMerges covering passthrough copy, linear merge, and SLERP.

---

Note

Medium Risk
Adds a new JSON architecture mapping and forces Pydantic model rebuilds for Configured*Architecture, which could affect architecture loading/validation across model types if the forward-reference resolution changes.

Overview
Adds first-class support for IBM Granite dense models by introducing a granite.json architecture definition (including optional bias tensors and tied lm_head handling) so merges no longer fall back to architecture inference.

Updates mergekit/architecture/base.py to import torch for PretrainedConfig forward-ref resolution and explicitly model_rebuild() the configured architecture Pydantic models.

Extends the test suite with a minimal GraniteForCausalLM fixture (make_picogranite) and new passthrough/linear/SLERP merge coverage for Granite models.

^{Reviewed by Cursor Bugbot for commit e1a7b40. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[Bhavyashah20]

I have read the CLA Document and I hereby sign the CLA

[cg123]

Thanks for the PR!