Check **rest in patterns for duplicates and alternations

[yuki-300]

(No Redmine ticket yet; I'll add the [Bug #NNNNN] prefix once it's filed.)

p_kwrest is the only binding rule in patterns that doesn't go through
error_duplicate_pattern_variable. p_variable, p_rest and the {a:}
shorthand in p_kw all do, but for **rest the name is handed straight to
assignable in new_hash_pattern_tail, so it never lands in pvtbl, is
never checked against in_alt_pattern, and never sets capture_in_pattern.
parse.y therefore accepts these while Prism rejects them:

in { a: x, **x }          # parse.y: runs, x ends up bound to the rest hash
                          # Prism:   duplicated variable name

in 1 | { a: 1, **rest }   # parse.y: Syntax OK under -c (compile.c still stops it at run time)
                          # Prism:   variable capture in alternative pattern

The duplicate case isn't a 4.0 regression, parse.y in 3.4.9 accepts it too
(Prism 3.4.9 already rejected it). The alternation case only shows up in 4.0
because that's where the parse-time check from [Feature #21572] came in;
under 3.4.9 both parsers let it through -c.

The fix is one line: call error_duplicate_pattern_variable from p_kwrest,
exactly like p_rest does. **nil is a different rule (p_kwnorest) and
doesn't bind anything, so in { a: 1, **nil } | 1 is still accepted.

Tests: assert_syntax_error for in {a: x, **x} next to the other
duplicated-variable cases in test_var_pattern, and assert_in_out_err(['-c'], ...)
for { a: 1, **rest } | 1 and 1 | { a: 1, **rest } in
test_alternative_pattern. test_pattern_matching.rb, test_syntax.rb and
test/ripper pass with RUN_OPTS=--parser=parse.y, and test_pattern_matching.rb
still passes under Prism.

This is kept separate from #1 (Bug #22310) because it turns code parse.y has
accepted since at least 3.4 into a SyntaxError, which isn't something I'd want
in a backport. It should go in after #1 though: once **rest sets
capture_in_pattern, the sibling leak that #1 fixes would otherwise start
rejecting in [ { **rest }, 1 | 2 ] as well. With both applied, parse.y and
Prism agree on every form I tried.