Fix display of nested/vlen compound and variable-length string array datasets

[brtnfld]

Summary

This resolves the customer-reported issue that #474 set out to fix (HDFView failing to display compound datasets containing nested/variable-length members), and additionally fixes a case #474 did not handle: arrays of variable-length strings, including arrays of arrays of them.

This branch is built by taking #474's (nested_seq_cmpd) legitimate compound/vlen display fixes and replacing its H5Dread_VLStrings/H5Dwrite_VLStrings workaround for array-of-varstr with a fix at the actual root cause, plus fixing the underlying table-view indexing bug that workaround was papering over. If this is merged, #474 can be closed - the commits from that branch are included here (with original authorship preserved) as the first 8 commits, so this is not a duplicate/independent implementation.

Root cause of the array-of-varstr bug

H5Datatype.allocateArray()'s ARRAY branch allocated a flat, narrowly-typed buffer (e.g. String[]) by delegating to the base type's own allocation. But per the JNI's documented buffer data model (translate_rbuf()/translate_wbuf() in HDF5's h5util.c), an ARRAY reads as an ArrayList of its elements whenever the base is itself variable-length - a flat buffer can't hold that. The resulting type mismatch throws ArrayStoreException inside the JNI's SetObjectArrayElement call, and translate_rbuf's own exception handling (CHECK_JNI_EXCEPTION(ENVONLY, JNI_TRUE)) silently discards it, leaving the buffer empty - no error surfaces anywhere.

I confirmed this empirically with a standalone array-of-varstr reproducer built against current HDF5 develop: the generic H5DreadVL/H5DwriteVL path returns empty strings on read and throws h5validate_wbuf: expected a java.util.ArrayList element on write, with the unmodified buffer-allocation code. Once the caller passes the documented Object[] shape instead, the same JNI call reads and writes correctly - no JNI changes needed.

Changes on top of #474's commits (net a code reduction)

• H5Datatype: new containsVlenOrVarStr(), mirroring the JNI's own recursive h5str_detect_vlen()/h5str_detect_vlen_str() (recurses through ARRAY bases and COMPOUND members to any depth). allocateArray's ARRAY branch now allocates Object[] instead of a flat buffer when the base needs it.
• H5ScalarDS: read/write dispatch now uses the same recursive check instead of the one-level-deep isArray() && base.isVLEN() test, which missed ARRAY of ARRAY of variable-length string. Fix display for nested/vlen cmpds #474's H5Dread_VLStrings/H5Dwrite_VLStrings special-casing for arrays is reverted back to plain isVarStr() - not needed once the buffer shape is correct.
• DataProviderFactory (ArrayDataProvider): the table view's flattened-offset indexing assumed one contiguous buffer spanning all rows, which cannot describe a List-per-row layout once nesting is involved (row 0 was pulling in row 1's data, row 1 went out of bounds, both showing *ERROR*). Added a container-based retrieval path that addresses nested List elements directly by index. The existing flat-buffer path is untouched and still used for plain numeric arrays.

Also two smaller fixes from review comments on #474 that apply independently of the array fix:

• DefaultBaseTableView: showUnsafeWriteNotice()'s one-dialog-per-second throttle is unnecessary - Tools.showInformation() opens a modal dialog, so the triggering event isn't re-entered until the user dismisses it.
• DefaultCompoundDSTableView: the cell-selection listener's early return keyed off unsafeForWrite was unrelated to write safety - it existed to dodge a baseIndexMap/expanded-column mismatch. Extracted the listener body into refineCellValuePreview() and replaced the flag check with a targeted try/catch around the actual failure mode.

Verification

Built HDF5 develop from source (Java bindings enabled) and tested against it end to end:

• Direct H5ScalarDS/H5CompoundDS probes: array-of-varstr, array-of-array-of-varstr, and compound-of-vlen-compound (the original reported bug) all read and write correctly.
• Actual HDFView table view (screenshots, clicked cells for full values): all three render correctly - e.g. the nested-array case shows [[r0e0a, r0e0b, r0e0c], [r0e1a, r0e1b, r0e1c]] exactly matching what was written, and the original compound-of-vlen file shows id/tags/nested columns with correct values - no more *ERROR* anywhere.
• mvn test -pl object: 163/163 pass, no regressions.

[brtnfld]

Closing in favor of contributing directly to #474 (see mattjala#2) rather than opening a parallel PR that duplicates the same commits already proposed there. #474 remains the canonical PR for this fix; my array-of-varstr fix and the remaining review-comment fixes will land as additional commits on that branch.