fix: Scope fontify buffers per session and skip redundant display updates

[dnouri]

Fontify buffers (" *pi-fontify:<lang>*") were globally shared by language. When two sessions wrote files in the same language, they corrupted each other's buffer — producing garbled syntax highlighting and triggering full O(n) re-fontification on every token instead of incremental O(1) appends. This caused multi-second freezes during write tool streaming in multi-session workflows.

Changes

• Each chat buffer now tracks its own fontify buffers via a buffer-local hash table (pi-coding-agent--fontify-buffers). Lookup is by hash table, not by buffer name. Buffers are killed on session end.

• Skip the delete+reinsert cycle in display-tool-streaming-text when the visible tail content is unchanged. Most LLM tokens extend a partial line not shown in the rolling preview, making ~90% of deltas no-ops for the buffer mutation path. fontify-sync still runs on every delta to keep the fontify buffer current.

• Remove handle-message-update, which accumulated text deltas via O(n²) concat into a field never read during streaming.

• Surface fontify-sync errors via message instead of swallowing silently with (condition-case nil ... (error nil)).

Benchmarks

Emacs 30.1, 500-delta write tool, xvfb real windows:

Metric	Before	After	Improvement
Two-session victim (mean)	36,145µs	1,537µs	23.5x faster
Two-session victim (p95)	84,656µs	2,753µs	30.7x faster
Single-session write (p50)	641µs	594µs	stable
Bash streaming (mean)	79µs	28µs	2.8x faster
Fontify buffer corruption	246 shrink events	0	eliminated
Fontify buffer leaks	confirmed	fixed	eliminated

Realistic partial-token benchmark (1,623 tokens, 171 lines of Python):

Metric	Value
Display updates	171 out of 1,623 deltas (89.5% skip rate)
Mean per-delta	1,063µs (includes 89.5% near-free no-ops)

Tests

• 483 unit tests pass
• 17 integration tests pass (real pi + Ollama)
• 13 GUI tests pass (real frames via xvfb)