[cmaloney]

Add working code generation for @vectorcall with **kwargs. The first commit adds support to libclinic converters to get keyword args from stack form rather than a dict. When there are no keyword arguments avoid creating a dictionary.

Exercise that code generation with xml.etree.ElementTree.Element. This adds a small behavior change in that non-string keyword arguments will raise TypeError rather than succeeding. Such keys can still be passed in using attrib. @text_signature keeps the documented (tag, /, attrib={}, **extra): attrib can still be passed by keyword, which the impl pops from extra.

Merging the keywords into the attributes now copies them with PyDict_Copy() instead of PyDict_New() plus PyDict_Update(), and pops attrib with _Py_ID(attrib) rather than creating a string per call; SubElement shares that helper.

pyperf microbenchmark (PGO+LTO, run serially):

Benchmark	base	Element
E('a', x='1')	116 ns	76.3 ns: 1.52x faster
E('a')	42.6 ns	28.2 ns: 1.51x faster
E('a', attrib=d)	122 ns	82.8 ns: 1.47x faster
E('a', d)	73.8 ns	53.6 ns: 1.38x faster
SE(E('p'), 'a', x='1')	161 ns	120 ns: 1.35x faster
E('a', d, y='2')	102 ns	80.3 ns: 1.27x faster
S('a')	49.4 ns	39.8 ns: 1.24x faster

Benchmark code:

import pyperf

runner = pyperf.Runner()
setup = ("from _elementtree import Element as E, SubElement as SE\n"
         "d = {'x': '1'}\n"
         "class S(E): pass")
for stmt in ["E('a')", "E('a', d)", "E('a', x='1')", "E('a', d, y='2')",
             "E('a', attrib=d)", "S('a')", "SE(E('p'), 'a', x='1')"]:
    runner.timeit(stmt, stmt, setup)

• Issue: Generate vectorcall code to parse arguments using Argument Clinic #87613