bug(scores): create_score sends numpy.bool_ as JSON true and Decimal as "<Decimal>" since v5.0.2 #2057

Describe the bug

#2006 (perf: build create_score body without Fern's dict(), in v5.0.2) replaced CreateScoreRequest(...) with a hand-built dict and kept only the bool/int → float coercion the model used to do. The model coerced every non-str value to float, so other numeric types now go out differently:

The server's score body takes z.union([z.string(), z.number()]) for value, so a JSON boolean is rejected. "<Decimal>" is accepted as a string, so the numeric score quietly turns into a categorical value "<Decimal>". This goes through every scoring path: create_score, span.score(), score_current_span/trace, and experiment item and run evaluators (Evaluation(value=...)), where numpy results are common.

To reproduce (no network, the score queue is captured)

import json, decimal, numpy as np
from langfuse import Langfuse
from langfuse._utils.serializer import EventSerializer

lf = Langfuse(public_key="pk-lf-test", secret_key="sk-lf-test", base_url="http://127.0.0.1:9")
sent = []
lf._resources.add_score_task = lambda score, **kw: sent.append(score)
for v in [np.bool_(True), decimal.Decimal("0.75")]:
    lf.create_score(name="accuracy", value=v, trace_id="0" * 31 + "1")
print([json.loads(json.dumps(s, cls=EventSerializer))["value"] for s in sent])
# main: [True, '<Decimal>']   v5.0.1: [1.0, 0.75]

Expected

Same payload as before #2006: any non-str value that float() accepts is sent as a float.

SDK: langfuse-python main ab861f1 (v5.0.4), Python 3.12.