fix(python): convert object-dtyped NumPy str/bytes arrays to pl.String/pl.Binary instead of pl.Object

[Wainberg]

Closes #10710 (which was originally about the conversion to NumPy, but this PR addresses the underlying issue better).

Before:

>>> pl.from_numpy(np.array(['a', 'b'], dtype=object))
shape: (2, 1)
┌──────────┐
│ column_0 │
│ ---      │
│ object   │
╞══════════╡
│ a        │
│ b        │
└──────────┘
>>> pl.from_numpy(np.array([np.nan, b'b'], dtype=object))
shape: (2, 1)
┌──────────┐
│ column_0 │
│ ---      │
│ object   │
╞══════════╡
│ nan      │
│ b'b'     │
└──────────┘

After:

>>> pl.from_numpy(np.array(['a', 'b'], dtype=object))
shape: (2, 1)
┌──────────┐
│ column_0 │
│ ---      │
│ str      │
╞══════════╡
│ a        │
│ b        │
└──────────┘
>>> pl.from_numpy(np.array([np.nan, b'b'], dtype=object))
shape: (2, 1)
┌───────────────┐
│ column_0      │
│ ---           │
│ binary        │
╞═══════════════╡
│ null          │
│ [binary data] │
└───────────────┘

This will lead to errors on mixed-type object arrays where the first element is a string or bytes. I'd originally considered something like:

first_type = type(values[0])
if first_type == str and all(type(v) == str for v in values[1:]):
    return PySeries.new_str
if first_type == bytes and all(type(v) == bytes for v in values[1:]):
    return PySeries.new_binary

but that would be much slower since you'd have to iterate in Python over the entire array. More importantly, it would be inconsistent with the pandas-to-polars conversion, which also just checks the type of the first (non-null) element.

[ritchie46]

Thanks!