Add support of kwargs to Backbone.from_preset and fix the dtype forwarding in Task.from_preset

[james77777778]

Currently, CausalLM.from_preset(..., dtype="bfloat16") has no effect because it doesn't forward dtype to the backbone.
This PR fixes that issue and also adds dtype support to Backbone.from_preset.

Additionally, I have updated the logic in load_serialized_object to support dtype when using DTypePolicyMap, ensuring that the pre-quantized preset will obey dtype:

import keras_nlp

llama_lm = keras_nlp.models.CausalLM.from_preset(
    "llama2_instruct_7b_en_int8", load_weights=False, dtype="bfloat16"
)
assert llama_lm.backbone.token_embedding.compute_dtype == "bfloat16"

[mattdangerw]

Looks good i think! just one small comment

[mattdangerw]

I believe the kwargs were supposed to ask as the config override directly (though it look like this broke at some point).

So if you wanted to set bert dropout, for example, you could do

model = keras_nlp.models.BertBackbone.from_preset("bert_base_en_uncased", dropout=0.5)

[james77777778]

Yeah, it is a bit weird to see config_overrides and I think there is a dangerous default value (config_overrides={}) for it in load_serialized_object.

I can try to fix it.

[james77777778]

@mattdangerw

I have updated Backbone.from_preset to support kwargs forwarding.
Now, this example works:

import keras_nlp


llama_lm = keras_nlp.models.CausalLM.from_preset(
    "llama2_instruct_7b_en_int8", load_weights=False, dtype="bfloat16"
)
assert llama_lm.backbone.token_embedding.compute_dtype == "bfloat16"

bert_backbone = keras_nlp.models.BertBackbone.from_preset(
    "bert_base_en_uncased", load_weights=False, dtype="bfloat16", dropout=0.5
)
assert bert_backbone.token_embedding.compute_dtype == "bfloat16"
assert bert_backbone.dropout == 0.5

[mattdangerw]

Thanks! This LGTM