Replace module hooking with tree-defined targeting

[Qubitium]

Even though existing code does not error, we should never ever hook/wrap a module that will never particpates in quantization. This doesn't make any sense. The fix is much more complicated and likely require for 2-3 PRs to refractor.

Currently add optional pin-point module targeting via tree of base.layers_modules_tree. This will play nicely into more generic multi-modal support as mm models embeds up to 3 separate models essentially into 1. The existing static-non-tree based config will be harder to support.

Tree syntax

# Full tree of quantizable modules
# `#` str will match any number: useful for layers and moe indexing.
# List[str] for serial linked nodes. List str are linear depth linked modules presented in a linear fashion with no divergence.
# Dict{str: List[str] | Dict | Tuple[str]} for diverging nodes where a node splits into multiple paths/nodes.
# Tuple(str) for final targeted modules/nodes: there are only strings representing the final targeted modules
layers_modules_tree = [
    "model",
    "layers",
    "#",
    {
        "self_attn": ("k_proj", "v_proj", "q_proj", "o_proj"),
        "mlp": ("up_proj", "gate_proj", "down_proj"),
    }
]

[Qubitium]

@Cecilwang While testing for conv1d/2d support I found deep issues/oversight with the current hooking targeting mechnaism so this Pr kind of spiraled out of control. But the code relevant to you is minimal. You just need to add the layer_modules_tree definition to your model for pinpoint control of hooks and check if the hooked_linear.py changes is usable for your model. I could not find a model that has nn.conv1d or nn.conv2d that is actually quantized. Let me know if you know any, other than mamba which I don't think we support yet.