Fix cspnet presets 3010

[pandeyshub-byte]

Description of the change

This PR adds 7 missing CSPNet presets to the preset map that were available on Kaggle but not accessible via from_preset().
While adding these, I found that the legacy weights (originally from KerasCV) use a configuration format that doesn't align with the current CSPNetBackbone constructor. To resolve this, I've:
Updated CSPNetBackbone with a from_config method to map legacy keys (like stackwise_channels and use_depthwise) to the current implementation.
Set appropriate architectural defaults (stage_type="cs3", expand_ratio=1.0) to match the Kaggle checkpoints.
Registered keras_cv>CSPDarkNetBackbone as a serializable alias so the loader recognizes the package name used in the original saved models.
I verified the fix by successfully loading all backbones and classifiers locally using the new preset strings.

Reference

Fixes #3010.
Kaggle models: https://www.kaggle.com/models/keras/cspnet

Colab Notebook

N/A (Preset addition and internal bug fix).

Checklist

• I have added all the necessary unit tests for my change.
• I have verified that my change does not break existing code and works with all backends (TensorFlow, JAX, and PyTorch).
• My PR is based on the latest changes of the main branch (if unsure, rebase the code).
• I have followed the Keras Hub Model contribution guidelines in making these changes.
• I have followed the Keras Hub API design guidelines in making these changes.
• I have signed the Contributor License Agreement.

[gemini-code-assist]

Code Review

This pull request adds several new preset configurations for CSP-DarkNet models to cspnet_presets.py and updates cspnet_backbone.py to support them. Key feedback includes correcting the indentation of the newly added from_config class method (which is currently defined at the module level), fixing an incorrect default value for expand_ratio in from_config, removing an unused import of backbone_presets, and correcting an implicit string concatenation bug in the docstring example for from_preset.

[gemini-code-assist]

This from_config method has two issues:

1. Indentation: It is defined at the module level, not as part of the CSPNetBackbone class. This will cause a NameError at runtime. It needs to be indented to be a class method.
2. expand_ratio default: The default value for expand_ratio is set to 0.5, which contradicts the PR description and the __init__ default of 1.0.

I've provided a suggestion below that fixes both the indentation and the default value for expand_ratio.

    @classmethod
    def from_config(cls, config):
        if "config" in config and isinstance(config["config"], dict):
            config = config["config"]

        config = dict(config)

        if "stackwise_channels" in config:
            config["stackwise_num_filters"] = config.pop("stackwise_channels")

        depths = config.get("stackwise_depth", [])
        num_stages = len(depths)

        if "stackwise_num_filters" not in config:
            config["stackwise_num_filters"] = [64, 128, 256, 512]

        if "stackwise_strides" not in config:
            config["stackwise_strides"] = [2] * num_stages

        if "stem_filters" not in config:
            filters = config.get("stackwise_num_filters")
            config["stem_filters"] = filters[0] // 2 if filters else 32

        config.setdefault("stem_kernel_size", 3)
        config.setdefault("stem_strides", 2)
        config.setdefault("block_type", "dark_block")
        config.setdefault("stage_type", "cs3")
        config.setdefault("expand_ratio", 1.0)
        config.setdefault("bottle_ratio", 0.5)

        valid_keys = [
            "stackwise_num_filters",
            "stackwise_depth",
            "stackwise_strides",
            "stem_filters",
            "stem_kernel_size",
            "stem_strides",
            "block_type",
            "stage_type",
            "expand_ratio",
            "bottle_ratio",
            "name",
            "trainable",
        ]

        return cls(**{k: v for k, v in config.items() if k in valid_keys})

[gemini-code-assist]

This import is unnecessary. The Backbone superclass automatically loads presets from the corresponding _presets.py file via its __init_subclass__ method. Please remove this unused import.