perf: Cache the rotation and skip redundant work in Transform2D

[spydon]

Description

Makes Transform2D cheaper on its hot paths without changing its API or behavior:

• The sine and cosine of the angle are cached and only recomputed when the angle actually changes. Previously every matrix recalculation called cos and sin, even when only the position, scale or offset had changed, which is by far the most common case (a component that moves every frame).
• The transformMatrix setter no longer throws away the matrix it was just given. It set _recalculate = false before updating the scale and position, whose listeners flipped it back to true, so the next read recomposed the matrix from the decomposed values. It now keeps the assigned matrix, and it updates the scale and position with one setValues call each instead of two separate x/y assignments, halving the number of notifications going out to their listeners.
• setFrom copies the fields directly (including the cached rotation) instead of going through the public setters, and Transform2D.copy is built on top of it.
• localToGlobal and globalToLocal read each matrix cell and point coordinate once instead of repeatedly, and hasReflection no longer goes through .sign.

A transform2d_benchmark.dart suite is added to packages/flame/benchmark so that these paths can be tracked going forward. Every case runs 1000 iterations per run.

Benchmarks

Times are per run (1000 transforms or point pairs), median of three rounds.

AOT (flutter run --release on an arm64 Android emulator, API 34, Apple Silicon host):

Case	Before	After	Change
position update + matrix read	250 µs	182 µs	-27%
angle update + matrix read	153 µs	158 µs	±0 (noise)
localToGlobal + globalToLocal	65 µs	68 µs	±0 (noise)
transformMatrix setter + matrix read	715 µs	504 µs	-30%
setFrom + matrix read	540 µs	398 µs	-26%

JIT (flutter test benchmark/transform2d_benchmark.dart, macOS, Apple Silicon):

Case	Before	After	Change
position update + matrix read	353 µs	300 µs	-15%
angle update + matrix read	201 µs	204 µs	±0 (noise)
localToGlobal + globalToLocal	49 µs	48 µs	±0 (noise)
transformMatrix setter + matrix read	896 µs	563 µs	-37%
setFrom + matrix read	662 µs	593 µs	-10%

The angle-update case is expected to stay flat since changing the angle still has to recompute the trigonometry; the point is that all other property changes no longer pay for it. The remaining cost of the position-update case is dominated by the ChangeNotifier round trip from NotifyingVector2 into Transform2D, which is out of scope here since the Vector2 implementation is about to change.

One thing worth noting for future work on this class: an earlier iteration of this PR inlined the recalculation check directly into localToGlobal/globalToLocal via a private _recalculateMatrix() method, and that made the point-conversion case about 50% slower under the JIT (76 µs vs 49 µs) while being neutral in AOT, presumably from the cold call site being inlined and bloating the hot methods. Going through the transformMatrix getter, as the code did before, keeps both compilers happy.

Checklist

• I have followed the Contributor Guide when preparing my PR.
• I have updated/added tests for ALL new/updated/fixed functionality.
• [-] I have updated/added relevant documentation in docs and added dartdoc comments with ///.
• [-] I have updated/added relevant examples in examples or docs.

Breaking Change?

• Yes, this PR is a breaking change.
• No, this PR is not a breaking change.

https://claude.ai/code/session_01H6f6E3vr3DM6iAWskqXpQT