[Impeller] Dispatch make-current lifecycle event only on success - #190655

Addresses flutter#190640

## Reproduction

Public deterministic reproduction:
[dfdgsdfg/mali-crash-app](https://github.com/dfdgsdfg/mali-crash-app).
The workload and the historical matrix table are documented in the
[README](https://github.com/dfdgsdfg/mali-crash-app#diagnostic-ab-matrix).

## Evidence

Confirmed facts from the diagnostic engine and the same production-like
scenario:

- The shared renderer deletes a texture name.
- The same numeric name is regenerated on the resource context.
- The upload path ends with `isTexture=false` for that name.
- The first isolated error is `FramebufferTexture2D`:
`GL_INVALID_OPERATION`, followed by a missing framebuffer attachment.

The [trace matrix
7268086803036215683](https://console.firebase.google.com/project/us-app-ea67d/testlab/histories/bh.d17e9c8f630d19e0/matrices/7268086803036215683),
[force-rebind pass matrix
7248171571680282013](https://console.firebase.google.com/project/us-app-ea67d/testlab/histories/bh.d17e9c8f630d19e0/matrices/7248171571680282013),
and [verify-3 matrix
9163761808229023214](https://console.firebase.google.com/project/us-app-ea67d/testlab/histories/bh.d17e9c8f630d19e0/matrices/9163761808229023214)
are the key Test Lab runs.

Inference: the trace and A/B results are consistent with a deleted
object remaining bound and Mali treating a rebind to the same numeric
name as no state transition. This is an inference about the driver
behavior, not a claim that the GLES implementation violates the
object-name specification.

## Change

The direct GLES buffer-to-texture upload path now performs one extra
zero bind immediately before binding the destination texture. The final
binding state is unchanged. The RGBA unit test uses deterministic
texture name `7` and preserves the ordered `0` then `7` binding
sequence.

This is a minimal workaround and does not supersede
[flutter#190655](flutter#190655): that PR
covers a different make-current failure path, while this reproduction
had a successful/current resource context. flutter#190655 was not directly
tested here.

## Physical validation

Using a Flutter 3.44.8-based diagnostic engine with the same app,
Scenario 9, and Mali device:

- Four affected diagnostic runs crashed after approximately 7–10
seconds.
- The force-rebind A/B passed 4/4 runs, each running for 606 seconds,
with no `GL_INVALID_OPERATION`, FBO failure, `isTexture=false`, or fatal
signal observations.

The latest-master patch received the host validation below; the physical
Test Lab binaries were built from the Flutter 3.44.8-based diagnostic
engine, not from this latest-master patch.

## Host validation

- `BlitCommandGLESTest.BlitCopyBufferToTextureCommandGLESRGBA`: 1/1
passed
- `BlitCommandGLESTest.*`: 11/11 passed
- `ProcTableGLES.*`: 5/5 passed
- `ReactorGLES.*`: 6 passed, 1 skipped because the host does not support
labelling
- `ImageDecoderNoGLTest.*`: 6/6 passed
- Host `impeller_unittests` and `ui_unittests` builds passed
- clang-format and `git diff --check` passed

## Scope and trade-offs

This is a narrow workaround validated for one deterministic
reproduction; it is not an architectural redesign or a general fix for
all GLES texture-binding paths. The current physical mechanism is
inferred from the trace and A/B results; this does not prove that every
Mali crash in flutter#190640 has the same cause.

The workaround applies only to `BlitCopyBufferToTextureCommandGLES`;
other GLES bind/upload paths remain unchanged. It adds one additional
`glBindTexture` per direct buffer-to-texture upload. Affected callers
include image decode, glyph atlas, small utility textures, and Flutter
GPU buffer-to-texture overwrite/copy paths.

The extra bind is likely small compared with texel transfer for normal
uploads, but it has not been benchmarked; repeated small per-frame
uploads may be sensitive. GL names come from the driver via
`glGenTextures`, and Flutter cannot prevent numeric name reuse.

Reactor-level name-reuse prevention is not a simple alternative: delayed
deletion and `glFinish` add memory and stall cost, and do not directly
force the observed same-name state transition.

Reviewer guidance on performance and scope, including broader workaround
plumbing if needed, is welcome.