Keep the cursor on control-expression keywords

[mmabdpr]

Fixes #3387

What changed

• Record cursor positions when printing if, then, elif, else if, match, match!, and with.
• Track else and if separately in ElseIfNode so extra spaces or line breaks in the source don't shift the cursor within either keyword.
• Formatting output, comment placement, and cursor fallback in whitespace remain unchanged.

Why

As noted in #3387:

1. genControlExpressionStartCore emitted keyword text directly without calling recordCursorNode.
2. Trivia.insertCursor only attached cursors to SingleTextNode, skipping the separate else and if nodes in ElseIfNode.

To fix this:

• Threaded recordCursorNode through genStart and genEnd in genControlExpressionStartCore.
• Replaced the anonymous object expressions in ElseIfNode with ElseIfKeywordNode and allowed insertCursor to attach cursors to them directly.

Design notes

• else and if tracked separately: Anchoring each token independently avoids cursor drift when else and if have extra spaces or sit on separate lines in the source.
• Direct type match vs. marker interface: I considered adding a CursorAnchor interface, but matched SingleTextNode and ElseIfKeywordNode directly in Trivia.fs to avoid adding a public abstraction for just two types. Happy to change this if you'd prefer an interface, or explore that in a follow-up PR.

Testing

Added 14 tests in CursorTests.fs covering short and multiline conditionals, elif, match, match!, comments around keywords, and else if with extra spacing or newlines. All 19 cursor tests pass.

[nojaf]

Thanks a lot for picking this up, @mmabdpr! Your analysis was spot on: the cursor recording that genControlExpressionStartCore skipped is exactly what broke if, then and match, and that part of your fix is in the final version.

When I wrote #3387, I already had some refactoring in mind that the issue didn't spell out. I should have been clearer about that up front, and it's why I went further on top of your branch instead of asking for more changes:

• ElseIfNode is gone. It used object expressions to reroute the trivia between else and if while it was being attached. IfKeywordNode.ElseIf now holds the two keywords as plain SingleTextNodes, and CodePrinter decides where the trivia ends up. The cursor then works on both keywords without any special node type. This makes ElseIfKeywordNode unnecessary.
• genControlExpressionStart no longer takes a Choice. match passes its keyword directly, and a new genIfStart handles if, elif and else if in one place.
• New genSingleTextNodeText helper. Code that writes a token's trivia separately from its text uses this helper, which also records the cursor. It turned up one more spot with the same bug, the closing bracket of a cramped list, which is fixed too.
• Only SingleTextNode stores a cursor now. It was the only node that ever held one, so the cursor members moved off Node.
• Tests rewritten in the existing style. I replaced the cursor tests with one keyword and one position per test, following the formatWithCursor |> assertCursor pattern already in CursorTests.fs. Sorry for the churn there. The project prefers tests that look like their neighbours over a new helper, even a well-built one.

None of this takes away from your work: your PR is what got this moving and showed exactly where the problem was. Thanks again, and I hope to see more from you!