fix: Geo::IP support — DynaLoader stubs, IPv6 Socket, nested eval STRING lexicals

[fglock]

Summary

Two fixes that together take ./jcpan -t Geo::IP from "configure fails immediately" to 8 of 8 test files passing.

---

Fix 1 — DynaLoader dl_* probe stubs + Socket IPv6 (commit 1)

Geo::IP's Makefile.PL calls DynaLoader::dl_findfile('GeoIP') to choose between XS and pure-Perl. We previously stubbed only bootstrap/boot_DynaLoader. Added no-op stubs for dl_findfile, dl_load_file, dl_find_symbol, dl_find_symbol_anywhere, dl_install_xsub, dl_undef_symbols, dl_error.

Also implemented Socket::pack_sockaddr_in6 / unpack_sockaddr_in6 (28-byte sockaddr_in6 layout: family, port, flowinfo, 16-byte addr, scope_id) and added them to @EXPORT, unblocking the PP-path IPv6 use Socket qw/ ... unpack_sockaddr_in6 /.

Fix 2 — Nested eval STRING visibility of outer my lexicals in named subs (commit 2)

Standard Perl lets a string-eval's compile-time scope include every my visible at the call site, including variables declared in an enclosing eval STRING. PerlOnJava's interpreter backend broke this for named subs defined inside a nested eval:

eval q{
    my $y = 99;
    eval q{ sub bar { return $y } 1 };
    bar();      # before: dies at parse-time or returns undef
                # after:  returns 99
};

Root cause: EvalStringHandler (interpreter backend) created a fresh, empty ScopedSymbolTable and received the caller's lexical environment only as a Map<name, register> used to build runtime capture arrays. The variable names never made it into the symbol table, so Variable.checkStrictVarsAtParseTime failed for named sub bodies, and SubroutineParser.handleNamedSub's JVM-path closure capture found nothing to capture.

Fix: Seed the parser's symbol table with one our entry per caller lexical, using the same "BEGIN package alias" trick that the JVM backend's RuntimeCode.evalStringHelper already uses:

1. Allocate a unique PerlOnJava::_BEGIN_<id> namespace per eval invocation.
2. Alias each captured my/state value into GlobalVariable.globalXxx under that namespace.
3. Seed addVariable(name, "our", seedPkg, null) so SubroutineParser.handleNamedSub takes its decl == "our" branch and looks values up via the alias.

Direct references inside the eval body are unaffected — they still go through BytecodeCompiler's parentRegistry-populated symbol table (register capture).

This mirrors exactly what the JVM backend has done correctly all along; the interpreter path was missing the alias step.

Geo::IP's remaining 2 test failures were entirely due to this: its v6 subs live in an inner eval <<'__IPV6__' that references outer-eval my @countries / @code3s / @names.

Design doc: dev/design/nested-eval-string-lexicals.md.

---

Verification

Check	Result
make (all unit tests)	green
make test-bundled-modules	green
New test src/test/resources/unit/eval_nested_lexicals.t (8 subtests)	8/8 pass, matches standard Perl
./jcpan -t Geo::IP	8/8 test files pass (was 0/8, then 6/8)
perl5_t/t/op/eval.t	159/173 pass (91.9%)
Reproducer from plan doc, direct/anon/named subs + 3-deep nesting	all match standard Perl

Files

File	Change
src/main/java/org/perlonjava/runtime/perlmodule/DynaLoader.java	Register dl_* probe no-op stubs.
src/main/java/org/perlonjava/runtime/perlmodule/Socket.java	Implement pack_sockaddr_in6 / unpack_sockaddr_in6.
src/main/perl/lib/Socket.pm	Export the two new functions.
src/main/perl/lib/DynaLoader.pm	Defensive .pm-level stubs (currently unreached — %INC is preset by Java — but useful if that changes).
src/main/java/org/perlonjava/backend/bytecode/EvalStringHandler.java	Seed symbol table with alias-package our entries; move capture computation before parse so seeding sees the final filtered set.
src/test/resources/unit/eval_nested_lexicals.t	New regression test.
dev/design/nested-eval-string-lexicals.md	Design doc.

Test plan

• make passes
• make test-bundled-modules passes
• ./jcpan -t Geo::IP passes all 8 test files
• New eval_nested_lexicals.t matches standard Perl
• No regression in perl5_t/t/op/eval.t

Generated with Devin