perf(import): read tags for a multi-file manual import preview in parallel

[jordanfelle]

DRAFT.

Problem

The Manual Import preview reads tags and duration (ffprobe or TagLib over network storage) one file at a time, which took about 340 s for a 3.5 GB, roughly 100-file audiobook on a cold cache (a 655-file audiobook took 577 s).

Fix

Warm the existing tag cache with four concurrent reads before the per-file loop; the loop then reads every file from the cache, so results are unchanged. Files that fail tag extraction are read twice.

Verification

• New SimpleImportDecisionMaker test (fails without the change).
• Core tests pass (3,021 on top of develop).
• Speedup not measured: there is no before/after timing on the same folder, so this stays a draft.

Review update:

• The warm-up only helps a cold folder: tags are also persisted in a database cache keyed by path, modification time and size, so a re-preview already skips tag reads. A 655-file audiobook preview that took ~577 s on a repeat run is therefore dominated by per-file matching (PerFileMatching runs a full staged match with FTS recall for every file; see the linked issue), not by tag reads. This PR stays a draft until a cold folder has been timed before and after.
• The tag cache is capped at 1,000 entries and clears itself when it overflows, so prefetching a bigger folder wiped the cache mid-prefetch. The prefetch is now capped at 800 files (test: multi_file_preview_should_not_prefetch_more_files_than_the_tag_cache_can_hold, failed with 887 reads against a cap of 800) (3,022 core tests pass).