Optimize SharedTransitionScope active transition tracking

Avoid iterating over the full sharedElements map whenever entries are
added or removed (in RememberObserver callbacks) or during state read
observations. Instead, track active transition status per entry on
SharedElementEntry with a custom setter, and maintain a counter
(activeTransitionsCount) on SharedTransitionScopeImpl for
isTransitionActive.

Furthermore, decouple state observations so each SharedBoundsNode only
observes its own entry's mutable properties (target, isEnabled,
isRunning, and deferred mutation state), completely eliminating
cross-node snapshot dependencies, and convert
_allEntries/_enabledEntries to plain lists.

Benchmark results (Pixel 6a, JIT disabled):
- Screen Instantiation (count=10): -19.7% to -29.2% faster (-179 us to
-217 us per initial composition), directly eliminating O(N^2) entry loop
overhead during onEntryAdded/onEntryRemoved.
- Screen Instantiation (count=1): -5.6% to -20.8% faster (-15 us to
-71 us per initial composition).
- Memory Allocations: 10% to 21% reduction in allocations across
composition and recomposition.
- Recomposition Toggle: -1.2% to -9.6% faster across single and
multi-element scenarios (-5 us to -16 us per toggle).
- Draw passes: -1.1% to -5.1% consistent improvements.
- Layout passes: Unchanged / within normal run-to-run noise (+/-0.1 us
to 5.0 us).

Test: new tests added
Fixes: 546240408
Change-Id: I97d72a9c89712d6c4d284f950a24913aa726da3b