Cache chat tab-line descriptors

[plux]

Description

ECA chat tab-line rendering rebuilt stable tab descriptors during redisplay.
That work repeated title formatting, activity checks, and session chat traversal for every redraw.

With many open chats, tab-line updates spent increasing time on data that only changes when chat state changes.
This PR caches stable tab descriptors per session.
It keeps per-redisplay selection state outside the cache.

Changes

• Cache stable chat tab-line descriptors per session.
• Keep selected state out of the stable cache so each buffer renders the correct active tab.
• Store chat activity state in cached descriptors so tab face selection avoids repeated buffer scans.
• Invalidate the tab-line cache when chat titles, loading state, pending approvals, chat registration, chat removal, chat exit, inline rollback, or session deletion changes.
• Add a session deletion hook so session-scoped caches clear before the session is removed.
• Add regression coverage for stale tab titles, selected state, active state, chat registration and removal, chat exit, and session deletion.
• Add benchmarks/eca-chat-tab-line-bench.el for focused tab-line timing and profiler runs.

Benchmarks

Measured benchmarks/eca-chat-tab-line-bench.el with synthetic chat buffers on Emacs 30.1.
Each chat had 20 turns.
Each result uses the median from 3 runs.

workload	chats	before	after	speedup
tabs-warm	10	140.74 us	25.02 us	5.63x
tabs-warm	50	638.53 us	78.44 us	8.14x
tab-face-all	10	96.48 us	24.74 us	3.90x
tab-face-all	50	545.82 us	106.76 us	5.11x

The normal warm tab-list path improved by 8.14x at 50 chats.
The tab face workload improved by 5.11x at 50 chats.

The forced rebuild path is slower after this change because that benchmark invalidates the cache before every call.
That row measures worst-case invalidation cost, not normal redisplay.
The intended path keeps stable tab descriptors cached between redisplays.

[ericdallo]

looks good, thanks