Rewrite nested hash collection pipelines without losing their types

[OskarEichler]

Summary

Pipeline::Insert descends only into Collect, missing the separate Collect::Hash subclass. Handle both and preserve concrete collection subclasses during recursive replacement/deletion.

Reproduction

require 'representable/json/hash'
a=->(*) { :old }; b=->(*) { :new }
result=Representable::Pipeline::Insert.call([Representable::Collect::Hash[a]], b, replace: a)
p result.first.first.equal?(b) # false before, true after

Verification

• Existing master suite on this isolated branch: 545 tests / 701 assertions, zero failures/errors, three existing skips on Ruby 3.2.11. Ruby 4.0.6 retains the same eight legacy inspection-format failures, zero errors and three skips. Existing tests/expectations are unchanged.
• 24 focused checks on Ruby 3.2.11/4.0.6 for this change. XML first-render uses a separate fresh-process check.
• Combined release-based verification with other focused fixes and attributed Use Setter in OverwriteOnNil if a :setter is defined #177/Hash value of null breaks parsing #264/Update MultiJSON usage to address deprecations #273 source backports: 544 tests / 700 assertions pass on Ruby 3.2.11, plus 4,410 focused checks and 36 boundary checks. Package retains 104 paths and unchanged compared metadata. All 36 runtime files compile on both Rubies; no newly introduced default Lint offenses.

Breaking changes / limitations

Requested nested hash functions now change; input pipelines remain unchanged and concrete collection types are retained.

No new/modified checked-in tests. Ruby 2.4–3.1, JRuby, Windows and Linux were not executed locally. Ruby >=2.4 and runtime dependency ranges remain unchanged. External test tooling supplies a legacy MiniTest alias and ruby-prof 1.7.x; no assertions were changed. No production verification is claimed.