Commit already-collapsed text in bulk in TreeBuilder::push_text

[nicoburns]

LLM Contributions: Code and this description were written by Devin (an LLM agent). The description needs rewriting by a human before this goes upstream.

Stacked on #62 (base is its branch).

In Collapse and PreserveBreaks modes, push_text stopped a text run at every collapsible character, so ordinary prose was committed one word at a time (each word going through flush_pending_whitespace, resolve_style_id, and commit_styled_text). A single space followed by content that needs no processing is already collapsed, so it can be committed as part of the run.

New collapsed_len(text, mode) replaces the find(is_collapsible || is_segment_break) that measured the text run:

scan bytes; stop at byte i if needs_processing(i) &&
    (bytes[i] != b' ' || i is last || needs_processing(i + 1))
where needs_processing = collapsible ASCII byte || CR || LF || LS/PS (E2 80 A8/A9)

So collapsing only runs at double spaces, tabs, newlines, LS/PS, and at the ends of the text. A trailing space still becomes pending whitespace, so behaviour across push_text calls and style spans is unchanged. The pending-whitespace, style and wrappability logic is untouched. #[inline(always)] is needed: without it, 1-char-per-push benchmarks were 12–30% slower.

Testing:

• Existing resolve::tree tests pass unchanged.
• A throwaway differential test (not included) compared against the old implementation on 300k random sequences of push_text, push_style_span, and pop_style_span. It mixed all four collapse modes, both wrap modes, and CR/LF/tab/FF/U+2028/2029/3000/lone E2 input. Text, style runs and style indices were identical. Deliberately breaking collapsed_len made the test fail.

Tango TreeBuilder push, compared paired against #62:

• Collapse, 10 chars per push, latin/arabic 400–8000 chars: 32–52% faster (e.g. latin 8000, 8 pushes/span: 102 µs → 53 µs).
• 20-char inputs: neutral to 19% faster.
• 1 char per push: neutral (0–5%).
• Japanese (no spaces, so nothing to merge): neutral to about 8% slower. For comparison, the Preserve rows, which run unchanged code, moved by ±4–12%.
• Preserve: unchanged code path.

Standalone 11.6 KB single-push benchmark, Collapse mode:

• ASCII prose: 39.7 µs → 11.0 µs.
• German prose: 37.3 µs → 10.9 µs.
• Indented HTML-like source: 36.9 µs → 14.5 µs.

Changelog: None

Link to Devin session: https://dioxus.staging.devinenterprise.com/sessions/c0f395eb38ed46d787feb525a9001029
Open in Devin Desktop: https://dioxus.staging.devinenterprise.com/desktop/session/c0f395eb38ed46d787feb525a9001029?variant=devin-insiders
Requested by: @nicoburns

[staging-devin-ai-integration]

I'll fix CI failures and address comments from users with write access that start with 'Devin'.

• Disable automatic comment, CI, and merge conflict monitoring