Fix unicode issues

[evertlammerts]

A Python str that UTF-8 cannot encode, such as "\ud800", failed on every value path with
RuntimeError: std::bad_cast: execute("SELECT ?", ["\ud800"]), executemany, sql(params=),
duckdb.Value, lists, dicts and maps, ConstantExpression, values, native UDF return values. In a
pandas or numpy object column it was worse: the scan handed the surrogate to an engine helper, which
raised an INTERNAL error and invalidated the whole database:

df = pd.DataFrame({"s": pd.Series(["\ud800"], dtype=object)})
con.sql("SELECT s FROM df").fetchall()
# INTERNAL Error: invalid code point detected in Utf8Proc::CodepointLength ...
con.sql("SELECT 42")
# FatalException: database has been invalidated

Both happened on 1.5.6 as well (there the first message was pybind11's "Unable to cast Python instance").

Root cause: nanobind's string caster clears the UnicodeEncodeError and reports a failed cast, so the
reason was lost at the first conversion; the object column scan decoded code points itself and never
checked for the surrogate range.

Changes:

• TransformPythonString converts a str with the same CPython call nanobind uses and, on a
  UnicodeEncodeError, raises InvalidInputException: Python Conversion Failure: cannot encode str as UTF-8, surrogate '\udXXX' at position N is not allowed. Used by the value converter, numpy <U enum
  entries, object category entries, and the read_csv na_values, delimiter and sep and write_csv
na_rep options. Valid strs take the same path as before.
• The object column scan (DecodePythonUnicode) computes UTF-8 lengths and encodes inline, raising the
  same error on a surrogate. It no longer calls the engine's per code point helpers, which also makes
  scans of non ASCII object columns without cached UTF-8 about 2x faster (measured on Release builds
  against the base; every other case is within noise).

Not changed: a str used as a name (a dict key becoming a field name, a named parameter name, SQL text)
still fails as before; duckdb.Value still checks nothing at construction. A str subclass such as
numpy.str_ with non ASCII text in an object column fails with an unrelated error, as on 1.5.6.