perf: template the kernel's `replace` on its callback (#15929)

This PR speeds up the kernel's expression traversals (`instantiate`,
`instantiate_rev`, `abstract`, `lift_loose_bvars`, `lower_loose_bvars`,
`instantiate_lparams`). `replace_tpl` is `replace` templated on its
callback, so the callback is inlined instead of being called through a
`std::function` at every visited node. It also takes a `skip` predicate:
subterms the callback would return unchanged are returned before the
cache lookup.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>