ui: stabilize the rendering of tool invocations

[zachwinter]

Overview

This PR addresses instabilities in the rendering of tool invocations in the llama-server UI.

This is a UI-only PR; no cpp core is touched.

Additional information

acd7a2d verifies a processing-state subscription leak as non-issue (lmk if we should keep this)
f25d4bf hardens tool call parsing; reduces header renders from 19 -> 1; language highlights 3 -> 1
9cdf7fa eradicates 24 blank frames during read-file parsing, reduces header renders from 2 -> 1
9a74042 MarkdownContent 11.50ms -> 2.18ms p95 per streamed token at 26KB
77fce09 edit-file; 16 all-deletion diff frames -> 0 per streamed edit
e1f38cf streaming stability contract across 7 parsers (flicker === gone)

Requirements

• I have read and agree with the contributing guidelines (YES)
• AI usage disclosure: yes; this is the second half of the work I did with Opus (followup to ui: reduce per-token render cost when streaming  #26053).

[ServeurpersoCom]

Can you drop acd7a2d? You measured the subscription leak at 0.005ms and concluded it wasn't worth fixing, so the 146 lines of fixture that prove it are dead weight in a suite that runs in a real browser on every PR. The rest of the perf harness is fine to keep, and streaming-stability.ts earns its lines.

One thing while you're in there: none of the perf fixtures ever close a code fence, they stream either plain paragraphs or a fence that stays open. That's how the endOffset issue got past the numbers.

This conflicts with #26097, which restructures parsePartialJsonArgs into scanPartialJson plus a memo cache. Which one goes first?

[ServeurpersoCom]

Let's just drop the perf suite entirely, it doesn't earn its place.

[zachwinter]

Let's just drop the perf suite entirely, it doesn't earn its place.

@ServeurpersoCom aye! would you like me to add to this PR's scope, or do this in a follow-up?

[ServeurpersoCom]

The scope just needs to be narrowly defined; browser rendering performance cannot be effectively tested via CI, it requires a human measured / manual testing to ensure there are no regressions / easy to review and no extraneous code.