perf: single-probe cache insert in replace_rec_fn

[Kha]

This PR replaces the find-then-insert pair in replace_rec_fn::apply with a single try_emplace on descent that reserves a slot for the result, then assigns into the slot on the way up via a reference that remains stable across std::unordered_map rehashing. This halves the hash work on the shared-but-uncached path and shaves ~4% off the wall-clock of leanchecker --fresh Init.Data.List.Lemmas.

[Kha]

!bench

[leanprover-radar]

Benchmark results for 0adccb2 against 82bb27f are in. Significant changes detected! @Kha

• 🟥 build//instructions: +27.0G (+0.22%)

Medium changes (5🟥)

• 🟥 elab/big_omega//instructions: +268.4M (+1.14%)
• 🟥 elab/big_omega_MT//instructions: +263.1M (+1.11%)
• 🟥 elab/big_struct_dep//instructions: +262.4M (+1.80%)
• 🟥 elab/grind_ring_5//instructions: +106.8M (+1.21%)
• 🟥 misc/leanchecker --fresh Init//instructions: +3.9G (+1.04%)

Small changes (10✅, 112🟥)

Too many entries to display here. View the full report on radar instead.

[mathlib-lean-pr-testing]

Mathlib CI status (docs):

• ❗ Mathlib CI can not be attempted yet, as the nightly-testing-2026-04-09 tag does not exist there yet. We will retry when you push more commits. If you rebase your branch onto nightly-with-mathlib, Mathlib CI should run now. You can force Mathlib CI using the force-mathlib-ci label. (2026-04-12 17:05:00)

[leanprover-bot]

Reference manual CI status:

• ❗ Reference manual CI can not be attempted yet, as the nightly-testing-2026-04-09 tag does not exist there yet. We will retry when you push more commits. If you rebase your branch onto nightly-with-manual, reference manual CI should run now. You can force reference manual CI using the force-manual-ci label. (2026-04-12 17:05:02)