锘縋artial backport of #3860 to 0.72-main (font-collection release only). Linked issue: #3535

Problem

On this branch TextMeasurer itself caches the OH_Drawing_FontCollection used for text measurement (there is no separate FontRegistry). When an app closes all RN surfaces while keeping the RNInstance alive, the cached collection stays resident on the native heap until instance teardown.

Scope

This is a reduced backport of #3860: on this branch the per-instance text measurement entry caches live in the global TextMeasureRegistry shared across instances, so trimming them from one instance's lifecycle would affect other instances. This change therefore only releases the per-instance cached font collection and leaves the global registry untouched.

Solution

• TextMeasurer: add releaseCachedFontCollection() dropping only the measurer's cached reference under its mutex; registered font bytes are kept, so the next getFontCollection() rebuilds the collection lazily (same behavior as the existing registerFont replacement path). The underlying OH_Drawing_FontCollection is destroyed via its registered OH_Drawing_DestroyFontCollection deleter once the last reference goes away.
• RNInstanceInternal: add releaseCachedFontCollection() forwarding to the measurer from the context container (find on the "textLayoutManagerDelegate" key), and call it in markSelfAboutToDestroyed().
• RNInstanceCAPI: call releaseCachedFontCollection() in destroySurface() when the last surface is removed. No createSurface() hook is needed: the collection is rebuilt lazily on the next measurement.

The host gtest additions of #3860 are not ported because this branch has no host gtest infrastructure; the release logic mirrors the FontRegistry::releaseCachedFontCollection() implementation covered by rnoh_font_test on main.

Impact

Internal lifecycle management only; no public API, symbol or behavior difference for surfaces that stay open. Typography consumers holding shared_ptr copies of the collection are unaffected - releasing only drops the measurer's cached reference.

Test

• Host gtest coverage for the same release logic exists on main via rnoh_font_test (PR #3860).
• Device A/B verification of the same source logic on later branches (5 runs each): baseline keeps roughly 120 MB/instance resident after all surfaces are closed while the instance stays warm; with the fix the measurement caches and the cached font collection are released at last-surface close and rebuilt on the next surface open, with identical measurement results.