fix(closure,HTML::Parser,open): unblock HTML-Tree 5.07 tests

[fglock]

Summary (WIP)

Fixes the three highest-impact bugs uncovered while investigating
./jcpan -t HTML::Element (HTML-Tree 5.07). Plan + remaining
work live in dev/modules/html_element.md.

Before this PR: 10 of 23 upstream test files fail.
After this PR: 3 of 23 fail. The remaining failures are:

• refloop.t — weak-ref destruction timing (Phase 4, being handled
  on another branch — explicitly out of scope here);
• whitespace.t — appears to hang on \xA0 input (separate issue);
• split.t — entity-handling diff (separate issue).

Bug 1 — For3Node continue block lost from closure capture

VariableCollectorVisitor.visit(For3Node) walked initialization,
condition, increment, and body — but not continueBlock.
The selective-capture optimisation in SubroutineParser used the
visitor's output to decide which outer lexicals to attach to a sub's
closure; anything referenced only inside a continue {} block
was filtered out. The lazy compiler then failed at first call with

Global symbol "$nillio" requires explicit package name

For1Node already traversed continueBlock correctly (lines 165-167);
only the For3Node visitor was wrong. Fix is one block.

Discovered via HTML::Element::look_down, which ends in:

while (defined($this = shift @pile)) { ... }
continue { unshift @pile, ..., @{ ... || $nillio }; }

where $nillio is a file-scoped my variable (see HTML/Element.pm
line 74 vs 2023).

Bug 2 — HTML::Parser tokens / tokenN / tokenpos argspecs

HTMLParser.java::buildArgs had no case for tokens, tokenpos,
or tokenN for N > 0. The default branch silently emitted an empty
scalar, which broke the default comment handler that
HTML/Parser.pm itself installs:

$self->handler(comment => sub {
    my ($self, $tokens) = @_;
    for (@$tokens) { $self->comment($_) }    # $tokens was ""
}, "self,tokens");

Implemented per HTML::Parser docs:

Event	tokens value
start	[tagname, k1, v1, k2, v2, ...] in attrseq order
end	[tagname]
text/dtext	[text]
comment	[comment_body]
declaration	[declaration_body]
process	[pi_body]

tokenN is tokens->[N]. tokenpos parallels tokens with
[0, 0] placeholder offsets (we don't track byte offsets — same
status as the existing offset / offset_end cases).

Bug 3 — literal-string filehandle name in open(...)

PrototypeArgs.handleTypeGlobArgument routed bareword filehandles
through FileHandle.parseBarewordHandle, which autovivifies
*main::FH and yields a typeglob reference. A constant-string first
argument fell through to the "Bare scalars" branch and was passed
in scalar context; open then tried to write the new IO into the
read-only string literal, producing
Modification of a read-only value attempted.

Now: when the first argument is a StringNode whose value is a
syntactically valid identifier (or Pkg::Identifier), treat it the
same as a bareword. open(my $fh, "<", $path) and other lvalue
paths are unaffected.

Discovered via HTML-Tree 5.07 t/oldparse.t:

open( "INFILE", "$TestInput" ) || die "$!";
binmode INFILE;
$HTML = <INFILE>;

Test plan

• make green on master (1m 11s)
• New regression tests:
  • src/test/resources/unit/closure.t — 2 cases for the
    continue-block capture bug
  • src/test/resources/unit/html_parser_tokens.t — 20 assertions
    for tokens / tokenN / tokenpos
  • src/test/resources/unit/open_string_filehandle.t — 5
    assertions for 2-arg / 3-arg / package-qualified / lvalue
• HTML-Tree 5.07 test suite, before -> after:
  | file | before | after |
  | --- | --- | --- |
  | attributes.t | 0/3 | 3/3 |
  | children.t | 0/3 | 3/3 |
  | comment.t | 0/1 | 1/1 |
  | construct_tree.t | 2/75 | 75/75 |
  | oldparse.t | 0/16 | 16/16 |
  | parse.t | 7/44 | 44/44 |
  | parsefile.t | 0/4 | 4/4 |
  | split.t | 1/444 | 417/444 (separate entity-handling issue) |
  | refloop.t | 3/8 (aborted) | 4/8 (Phase 4, out of scope) |
  | whitespace.t | 0/6 (aborted) | 4/6 then hangs on \xA0 |

Out of scope / follow-ups

• Phase 4 (-weak refloop) — being addressed on a separate
  branch; do NOT implement here.
• split.t entity diff — output writes & literals
  through some paths instead of decoded entities. Separate PR.
• whitespace.t hang on \xA0 — likely an HTML::TreeBuilder
  whitespace-folding loop; needs its own investigation.

Generated with Devin