Fix InexactError in peek_behind_pos when skipping nested trivia nodes

[KristofferC]

Explanation of bug written below by Claude :robot. Fixes JuliaLang/julia#60084

---

When parsing certain incomplete keywords like "do", the parser would crash
with InexactError: trunc(UInt32, -1) instead of returning a ParseError.

Root cause:

The peek_behind_pos function walks backwards through the flat output array,
subtracting byte_spans to calculate byte positions. However, when a non-terminal
trivia node contains child nodes, both the parent and children exist in the
output array with overlapping byte_spans. The old code would subtract the
byte_span of both the parent and its children, causing byte_idx to go negative.

Example with input "do":

The parser creates this output structure:
[1] TOMBSTONE (byte_span=0)
[2] "do" token (byte_span=2, is_trivia=true, covers bytes 1-2)
[3] error node (byte_span=2, is_trivia=true, node_span=1, covers bytes 1-2)
└── contains node [2] as a child

When peek_behind_pos walks backwards from next_byte=3:
Old behavior:
- Start: byte_idx = 3, node_idx = 3
- Skip error node [3]: byte_idx = 3 - 2 = 1, node_idx = 2
- Skip "do" node [2]: byte_idx = 1 - 2 = -1 ❌ (tries to convert to UInt32)

The problem: Both nodes cover the same bytes (1-2), so subtracting both
spans double-counts the same 2 bytes.

Solution:

When skipping a non-terminal node, also skip all its children without
subtracting their byte_spans, since the parent's byte_span already includes
them:

if is_non_terminal(node)
node_idx -= (1 + node.node_span) # Skip the node + all its children
else
node_idx -= 1
end

New behavior:

• Start: byte_idx = 3, node_idx = 3
• Skip error node [3]: byte_idx = 3 - 2 = 1, node_idx = 3 - (1 + 1) = 1
• Stop at TOMBSTONE [1] (not trivia)
• Return: byte_idx = 1, node_idx = 1 ✓

Why "do" was unique:

The "do" keyword is the only Julia keyword that:

1. Gets marked as TRIVIA when appearing standalone (invalid syntax)
2. Has an error node emitted that wraps it (also marked as TRIVIA)
3. Continues parsing afterward, eventually calling peek_behind

Other incomplete keywords either parse successfully or fail earlier before
reaching the code path that calls peek_behind.

[Keno]

The fix seems non-controversial, although I'm not sure that non-terminal trivia are that good an idea, but that's a bit of a separate question. Given that we have them right now, this is the right fix.