perf(parser): cut redundant NNTP reads and cap PAR2 scan

[javi11]

Summary

Three optimizations to internal/importer/parser that together cut parse-time NNTP reads by ~50–70% on typical multi-file releases.

1. Share yEnc standardPartSize across the NZB

normalizeSegmentSizesWithYenc previously fetched the second and last segment of every multi-segment file to learn the yEnc PartSize. Every file produced by the same encoder shares the same middle-segment size, so we now fetch it once per NZB via pickRepresentativeMiddleSegment and reuse it. Per-file normalization only fetches the last segment.

• On a 50-file NZB: ~49 fewer priority-lane reads.
• Fallback: if the representative fetch fails, normalization falls back to the old per-file behavior.

2. Defer 16KB first-segment completion behind PAR2 presence

The fan-out loop in fetchAllFirstSegments that reads extra segments to reach 16KB exists only for PAR2 Hash16k MD5 matching. It now runs only when hasPar2IndexCandidate confirms the NZB actually contains a PAR2 index, and it skips obvious sidecars (.nfo, .txt, .srt, .sub, .jpg, .jpeg, .png, .nzb, .sfv, .md5).

• NZBs without PAR2 (Stremio-fetched releases, obfuscated one-file NZBs) pay zero extra reads here.

3. Cap PAR2 scan timeout and short-circuit on FileDesc completion

par2/descriptor.go previously used 30s × segments (up to 2.5 min for a 5-segment index) and read up to 1000 packets. Now:

• Timeout capped at 90s.
• Scan breaks after 50 non-FileDesc packets following the last descriptor — healthy index files exit as soon as the FileDesc section ends.

4. Parallelize PAR2 extraction and representative yEnc fetch

par2.GetFileDescriptors and the one-shot representative yEnc fetch are independent. They now run together in an errgroup, so wall time is max(par2, yenc) instead of par2 + yenc.

Notes for reviewers

• No change to ConnectionPool / BodyAsync contracts.
• No change to notFoundIDs semantics.
• Output of ParsedNzb should be byte-identical for existing NZBs — the only behavioral change is when / whether certain reads happen, not which data lands in the parsed result.
• Correctness anchor for change feat: add documentation and multiple fix #2: fileinfo.getFileInfo already handles the case where the padded MD5 doesn't match any descriptor (it's the normal path for non-PAR2 NZBs), so skipping 16KB completion is safe.

Test plan

• go build ./...
• go vet ./internal/importer/parser/...
• go test -race ./internal/importer/parser/... — all existing tests pass
• End-to-end smoke: import a known-good 50-file RAR NZB locally and confirm ParsedNzb matches byte-for-byte vs. pre-change output, and that parse time drops meaningfully via progressTracker timings
• End-to-end smoke: import a non-PAR2 single-file NZB and confirm no additional-segment reads are issued