feat(pkg-py): Replace pandas with narwhals

[cpsievert]

This PR removes pandas as a required dependency and replaces it with narwhals, a lightweight DataFrame abstraction layer that supports both pandas and polars backends. Users can now choose their preferred DataFrame library.

Motivation

• Reduced dependencies: pandas is a heavy dependency; users who prefer polars shouldn't need to install pandas
• Flexibility: Some users prefer polars for its performance characteristics
• Future-proofing: narwhals provides a stable API that works across DataFrame libraries

Changes

Dependencies

• Removed pandas from required dependencies
• Added pandas and polars as optional dependencies
• Install with pip install querychat[pandas] or pip install querychat[polars]

API Changes (Breaking)

• execute_query(), get_data(), and df() now return narwhals DataFrames instead of pandas DataFrames
• Call .to_native() on returned DataFrames to get the underlying pandas/polars DataFrame

Internal Changes

• New _df_compat.py module handles backend selection (prefers polars when available)
• df_to_html() generates HTML directly without pandas dependency
• DataFrameSource accepts pandas, polars, or narwhals DataFrames

Tests

• Added test_df_compat.py for the compatibility layer
• Added test_dataframe_source.py with comprehensive DataFrameSource tests
• Polars-specific tests are skipped when polars isn't installed

Migration Guide

# Before
df = querychat.df()  # pandas DataFrame
df.head()

# After
df = querychat.df()  # narwhals DataFrame
df.head()            # still works - narwhals has similar API

# If you need a native pandas/polars DataFrame:
native_df = querychat.df().to_native()

[cpsievert]

I'm pretty sure it's going to make sense for us to have a separate LazyFrameSource, which I'll do in a follow up PR (before the next release). The benefit being that we can be more lazy about computation, and possibly have .df() also return a LazyFrame in that scenario

[cpsievert]

Note that downstream calculation of ranges and unique values wasn't working properly because they were based on the first 10 rows -- I'll address this when doing the new LazyFrameSource implementation