[WIP]support automatic mixed bits assignment

[wenhuach21]

• support 16 bits in options
• support gguf
• support mixed data type, algorithm could mixed, exporting could not
• gguf group_size fallback
• check gguf scale_dtype
• layer_config ut
• support FP8 model
• AutoScheme could also be patched by group_size, etc, fill with default value
• gguf could not be mixed with other formats
• support non_str dataset
• support shared_layers
• support naive methods
• test_large_model

[wenhuach21]

@xin3he @n1ck-guo
please have a review of this function first

def compute_layer_bits(
    layer: torch.nn.Module,
    ignore_scale_zp_bits: bool = False,
) -> tuple[int, float]:
    """Compute total and average bitwidth for a single quantized layer.
    Args:
        layer: A PyTorch layer with quantization attributes.
        ignore_scale_zp_bits: Whether to ignore scale/zero-point overhead.
    Returns:
        A tuple (total_bits, avg_bits) representing bit usage.
    """
    weight = layer.weight
    n_param = weight.numel()
    weight_bits = getattr(layer, "bits", 16)
    group_size = getattr(layer, "group_size", 128)
    super_group_size = getattr(layer, "super_group_size", None)
    super_weight_bits = getattr(layer, "super_bits", None)

    # Unquantized layer or ignoring scale/zp overhead
    if weight_bits >= 16 or ignore_scale_zp_bits:
        if super_weight_bits is not None:  # reset gguf 16 bits to 32 bits
            return 32 * n_param, 32
        return weight_bits * n_param, 16.0

    in_features, out_features = get_layer_features(layer)

    # Determine number of groups based on group size
    if group_size > 0:
        n_group = out_features * (in_features + group_size - 1) // group_size
    elif group_size == 0:
        n_group = 1
    elif group_size == -1:
        n_group = out_features
    else:
        raise ValueError(f"Invalid group_size {group_size}")

    # Compute auxiliary bits (scales, zero-points, or double quantization)
    aux_total_bits = 0
    if not super_group_size:
        scale_bits = 16
        zp_bits = weight_bits
        aux_total_bits = n_group * (scale_bits + zp_bits)
    else:
        aux_total_bits += n_group * super_weight_bits * 2
        n_super_group = (n_group + super_group_size - 1) // super_group_size
        aux_total_bits += n_super_group * 32 * 2  # 32-bit scale and min_v

    total_bits = weight_bits * n_param + aux_total_bits
    avg_bits = total_bits / n_param
    return total_bits, avg_bits