Fix: Discovery review selection lost due to duplicate tree-data field

[jpedane]

Symptom: In Discovery > Task list > Review, selecting devices/modules and pressing OK always returns "No changes. Re-Scheduled" and no agents are created, even though valid devices were selected.

Root cause chain:

1. load_modal() in pandora_ui.js executes every inline <script> in a modal's AJAX response TWICE. Its success handler regex-extracts the script bodies and runs them through $.globalEval() after calling settings.target.html(data) but jQuery's .html() already executes embedded scripts during insertion (it only strips the <script> tags from the resulting DOM, which is what the misleading comment above the globalEval loop refers to). Every modal inline script therefore runs twice per open, deterministically.

2. The review tree script emitted by HTML::printTree() appends a hidden into the tree container each time it runs, with no existence check. Because of (1) it runs twice, so two same-named inputs end up in the form. The tree widget's onChange writes the selection via $("#tree-data-tree"), which matches only the FIRST input; the second stays empty forever.

3. On OK, load_modal's serializer walks '#review :input' and appends every matching element to the multipart FormData individually, so both fields are submitted populated first, empty second.

4. PHP keeps only the LAST value for a repeated POST field name, so parseTaskReview() reads get_parameter('tree-data-tree', '') as an empty string, matches nothing, and answers "No changes. Re-Scheduled".

Fix (this commit): make printTree() idempotent only append the hidden input if it does not already exist. With the guard, both script executions share one input, the selection survives serialization, and parseTaskReview() receives it. Also fixes an adjacent bug on the same lines where the tree root selector was emitted unquoted (el: $(tree) instead of el: $("#tree")), which only worked by accident via the browser's named-element window globals.

Not fixed here (follow-up candidate): the underlying double execution in load_modal() still runs every modal inline script twice; for this modal it still renders the device tree twice into the container (cosmetic after this fix). Removing the redundant globalEval loop in pandora_ui.js would fix that globally but affects every modal in the console, so it deserves its own change and testing.

Verified end-to-end with jsdom + the exact jquery-3.6.0.min.js and simTree.js shipped in this repo, replaying load_modal's injection and serialization code verbatim against the real printTree() output from PHP CLI:

BEFORE: 2 tree-data-tree inputs in DOM, 2 fields serialized, second
empty -> PHP receives '' -> "No changes. Re-Scheduled"
AFTER: 1 input, 1 field serialized with the selection JSON ->
agents/modules scheduled for creation

php -l passes on the modified file.