[quantization] Add builder helper for PTQConfig

[mhs4670go]

Motivation

Constructing PTQConfig.overrides for LLM models is currently verbose and error-prone. Users must manually define deeply nested module paths (e.g., model.layers.{i}.self_attn.q_proj) even for very common patterns. This leads to:

• Poor readability in example scripts
• High maintenance cost when model structures differ slightly
• Increased risk of silent misconfiguration (e.g., incorrect key paths)

We should provide a helper that encapsulates these patterns while remaining extensible across model families.

Proposal

Introduce a model-aware builder:

build_llm_ptq_config(...)

Key idea:

• Public API is model-agnostic
• Internally dispatches to model-specific mapping logic

Example:

cfg = build_llm_ptq_config(
    model_type="llama",
    num_hidden_layers=len(model.model.layers),
    wrapper_variant="prefill",
    activation_dtype=DType.int(16),
    default_qscheme=QScheme.PER_TENSOR_SYMM,
    linear_weight_bits=4,
    embedding_weight_bits=8,
    lm_head_weight_bits=4,
)

[stamalakhov]

I believe hiding explicit construction of PTQConfig (currently verbose) construction into helper is a good idea.

[mhs4670go]

Also, we need to pair dtype and qscheme. For instance, uint8 and ASYMM don't have to be given simultaneously. Let's automatically infer qscheme by given dtypes in the next issue.

qcfg.overrides["model"]["rotate_embedding"] = {"weight": {"dtype": DType.uint(8)}, "qscheme": QScheme.PER_TENSOR_ASYMM}
qcfg.overrides["rotate_lm_head"] = {"weight": {"dtype": DType.uint(8)}, "qscheme": QScheme.PER_TENSOR_ASYMM}
qcfg.overrides["lm_head"] = {"weight": {"dtype": DType.uint(8)}, "qscheme": QScheme.PER_TENSOR_ASYMM}
for idx in range(len(q_m.model.layers)):
    qcfg.overrides["model"]["layers"][str(idx)]["self_attn"]["q_proj"]["act_in"] = {"dtype": DType.uint(8), "qscheme": QScheme.PER_TENSOR_ASYMM}
    qcfg.overrides["model"]["layers"][str(idx)]["self_attn"]["k_proj"]["act_in"] = {"dtype": DType.uint(8), "qscheme": QScheme.PER_TENSOR_ASYMM}
    qcfg.overrides["model"]["layers"][str(idx)]["self_attn"]["v_proj"]["act_in"] = {"dtype": DType.uint(8), "qscheme": QScheme.PER_TENSOR_ASYMM}
    qcfg.overrides["model"]["layers"][str(idx)]["self_attn"]["o_proj"]["act_in"] = {"dtype": DType.uint(8), "qscheme": QScheme.PER_TENSOR_ASYMM}
    qcfg.overrides["model"]["layers"][str(idx)]["self_attn"]["o_proj"]["weight"] = {"dtype": DType.uint(8)}
    qcfg.overrides["model"]["layers"][str(idx)]["mlp"]["up_proj"]["act_in"] = {"dtype": DType.uint(8), "qscheme": QScheme.PER_TENSOR_ASYMM}
    qcfg.overrides["model"]["layers"][str(idx)]["mlp"]["gate_proj"]["act_in"] = {"dtype": DType.uint(8), "qscheme": QScheme.PER_TENSOR_ASYMM}
    qcfg.overrides["model"]["layers"][str(idx)]["mlp"]["down_proj"]["weight"] = {"dtype": DType.uint(8)}