Resolve link topics against a per-run alias index

[hadley]

Profiling roxygenise() on a link-heavy package (testthat) showed ~50% of the roxygen work inside link/topic resolution: has_topic() answers every query with a utils::help() call (~1.5ms), and find_package_lookup() maps it over all dependencies plus base packages, so each unique linked topic costs ~50 help-index searches. The per-topic cache added in #1724 only deduplicates repeated topics — every miss still pays the full scan.

This PR replaces the help() calls with a cached topic index per package:

• For installed packages, read the help/aliases.rds index that ships with the package (once per package per session — installed packages don't change between runs).
• For packages loaded from source — the package being documented, or a dependency loaded with pkgload::load_all() — reuse pkgload::dev_topic_index(), the alias index that dev_help() already maintains. (pkgload cached this before too, but the old path reached it through help() → shim → dev_topic_parse()/normalizePath() on every call, and the cache never applied to the ~50 installed dependencies, which were the bigger cost.)
• Topics are stored as a hashed environment so each has_topic() is a single env_has(); %in% on the alias vector rebuilt a hash table per membership test, which the R6 and Rd-inheritance paths (no per-topic cache in front of them) paid heavily.
• pkg_deps() is also cached per run — it re-read and re-parsed DESCRIPTION for every unresolved topic.
• At the start of each run, find_package_cache_reset() evicts only the documented package's entry (from our cache and pkgload's) — its topics are the only ones that change between runs. Failed lookups are never cached, so installing a missing dependency mid-session is picked up immediately.

check_topic() (validation of explicitly-qualified links) and the Rd-inheritance and R6 code paths that use has_topic() get the same speedup for free. Re-export attribution (find_source()), the ambiguous-topic warning, and all warning messages are unchanged.

Benchmark on a quiet machine (medians of 4 adjacent steady-state runs, spreads ±0.04s): roxygenise() on testthat goes from 3.00s to 2.15s (1.4x), and to 2.05s (1.5x) combined with the markdown tokenizer rewrite in #1901. Note ~0.9s of every run is fixed load_all() + Rd-writing overhead, so the roxygen-work portion roughly halves. Link resolution falls from ~50% of the profile to under 8%.

🤖 Generated with Claude Code

[hadley]

Added the last two low-hanging items the combined-branch profile surfaced (8.2% + 5.0% of samples):

• pkg_deps() re-read and re-parsed DESCRIPTION for every unresolved topic — now cached per package dir, reset each run.
• has_topic() used %in% on the alias vector, rebuilding a hash table per membership test — pkg_topics() now caches a hashed environment, so each lookup is a single env_has(). This mostly benefits the R6 and Rd-inheritance paths, which call has_topic() directly without the link path's per-topic cache.

With these, link/topic resolution is essentially free; the remaining profile is the markdown XML walk (~23%, a structural project) and diffuse Rd assembly.