perf(crawler): parse HTML once per page in FessXpathTransformer

[marevol]

Summary

Parse each crawled HTML page's DOM once instead of twice. FessXpathTransformer previously parsed the response body into a DOM in storeData (field/anchor extraction) and then the inherited HtmlTransformer.storeChildUrls parsed the same body a second time for child-URL discovery.

Dependency

Requires codelibs/fess-crawler#177 (the storeChildUrls(ResponseData, ResultData, Document) overload). Merge that first; this PR builds against fess-crawler 15.8.0-SNAPSHOT including it.

Changes

• Stash the Document parsed in storeData into a ThreadLocal (the transformer is a shared singleton across crawler threads, so an instance field would not be thread-safe).
• Override the two-argument storeChildUrls to reuse that Document via the new base three-argument overload instead of re-parsing. Its exception handling mirrors the base method exactly — and since ChildUrlsException extends CrawlerSystemException, the canonical/noindex redirect mechanism is preserved with its original type.
• Wrap transform() in a try/finally that clears the ThreadLocal, so a pooled crawler thread never leaks or reuses a stale DOM. A defensive null check falls back to the base (re-parsing) method.
• Clone-before-prune fix: the field-rule loop's default branch prunes a deep clone instead of the live node.

Behavior notes

Because the DOM is now shared between field extraction and child-URL/anchor discovery:

• Clone-before-prune prevents links nested under pruned subtrees (e.g. a <nav> inside a pruned //H1///H2///H3) from being silently dropped from child-URL discovery. The indexed field value (getTextContent() of the pruned clone) is byte-identical to before.
• Child-URL/anchor extraction now runs on the storeData parse (Fess-detected charset via InputSource.setEncoding + UTF-8 BOM strip) instead of a separate neko auto-detected parse — a consistency improvement, converging on the charset already used for content/anchor extraction.
• The indexed anchor field can now also retain links under pruned-tag subtrees nested in headings.

Testing

• test_transform_linkUnderPrunedHeading_isDiscovered — a link nested under a pruned <nav> inside a pruned heading is still discovered. Verified to FAIL if the clone-before-prune fix is reverted to in-place pruning (the link is dropped), i.e. it genuinely guards the regression.
• test_transform_parsesBodyOnce — asserts the response body is read exactly 3 times per transform() (charset sniff + storeData parse + document-cache read); there is no 4th read, confirming child-URL discovery no longer re-parses.
• test_transform_normalPage_fieldsAndChildUrlsUnaffected — ordinary pages extract identical fields and child URLs.
• FessXpathTransformerTest: 65 tests pass; mvn formatter:format / license:format clean.