feat: add pandas df accessor

[shreyashankar]

Adds a pandas DataFrame accessor (.semantic) that provides LLM-powered operations through DocETL's engine. This integration is inspired by the LOTUS paper (Patel et al. 2024) and enables:

• Semantic mapping with LLMs (df.semantic.map())
• Semantic filtering (df.semantic.filter())
• Fuzzy merging of DataFrames (df.semantic.merge())
• Semantic aggregation with optional fuzzy matching (df.semantic.agg())
• Automatic cost tracking and operation history

Example usage:

import pandas as pd
from docetl import SemanticAccessor

df = pd.DataFrame({"text": ["Apple released iPhone 15", "Google launches Pixel 8"]})
df.semantic.set_config(default_model="gpt-4o-mini")
result = df.semantic.map(
  prompt="Extract company and product from: {{input.text}}",
  output_schema={"company": "str", "product": "str"}
)

# result is a df with 2 new cols: company and product

Note: While individual operations are optimized internally, pipelines created through the pandas interface cannot be optimized as a whole. For pipeline-level optimizations, use the YAML or Python API interfaces.