fix(eval STRING): preserve 'our' decls so caller-side local() is visible

[fglock]

Summary

Fixes eval STRING so that local applied to outer our-declared
package variables is visible inside the eval body.

Minimal reproducer that broke before this patch:

our $x = 1;
sub test { local $x = 42; eval q{print "eval x=$x\n"}; }
test();   # was printing 1, should (and now does) print 42

The same issue affected local @our_array / local %our_hash seen
from inside eval STRING, including element access like $our_hash{k}.

Root cause

The bytecode-interpreter eval path (default on PerlOnJava) builds an
adjustedRegistry of captured variables in RuntimeCode /
EvalStringHandler and hands it to BytecodeCompiler, which seeded
every captured variable in its symbol table with decl="my".
compileVariableReference then took the register path for our vars
instead of emitting LOAD_GLOBAL_SCALAR, so the eval body read the
scalar that was in the global slot at eval-entry time — not the
currently-localized one.

Changes

• RuntimeCode.evalStringWithInterpreter now builds a parallel
  adjustedDecls map from the captured symbol table and passes it to
  BytecodeCompiler so our stays our.
• BytecodeCompiler gets a new constructor overload that accepts
  parentDecls and preserves the outer decl when seeding its symbol
  table.
• handleArrayElementAccess and handleHashElementAccess now check
  isOurVariable and route our aggregates through
  LOAD_GLOBAL_ARRAY / LOAD_GLOBAL_HASH so element reads see the
  current (possibly localized) slot.
• EvalStringHandler deliberately does not propagate
  adjustedDecls: nested evals seed outer my vars as our in a
  synthetic package purely for the parser, and at runtime those must
  stay register-bound. A comment explains this.

Observable impact

jcpan -t TimeDate goes from 340/1215 failing subtests across 10
files to 8/1215 in 1 file (t/lang-encoding.t, a pre-existing
utf8::is_utf8 issue unrelated to eval/local). The original
Date::Parse::str2time failures ("Fri Dec 17 00:00:00 1901 GMT",
"Wed, 16 Jun 94 07:29:35 CST", …) were all caused by this bug:
Date::Parse builds its regex closures via eval "$strptime" inside
local($day_ref,$mon_ref,$suf_ref), and the eval was seeing the
pre-local (undef) scalars instead of the freshly-localized ones.

Test plan

• make (all unit tests pass, including
  unit/eval_nested_lexicals.t which exercises the
  EvalStringHandler path we intentionally left alone)
• make test-bundled-modules
• jcpan -t TimeDate — 1/28 test files fail, 8/1215 subtests
  fail (down from 10/28 and 340/1215)

Generated with Devin