gpu: use 1-byte pixel row alignment for texture upload and read back

[alepacheco]

The image buffers that uhdr_opengl_ctxt uploads with glTexImage2D() are tightly packed: each row is w * bytes per pixel long. OpenGL ES reads them with the default GL_UNPACK_ALIGNMENT of 4, which assumes every row starts on a 4-byte boundary. For the 8-bit YCbCr formats (uploaded as GL_R8) and for RGB888, a width whose row size is not a multiple of 4 makes the driver read past the end of the buffer and shears the texture. Only 8bppYCbCr400 already set the alignment to 1 around its upload and read back.

With GPU acceleration enabled this is reached by decoding, for example, a 4:4:4 base JPEG whose width is not a multiple of 4: applyGainMapGLES() uploads the contiguous YUV planes and the driver reads beyond the allocation.

This PR sets GL_UNPACK_ALIGNMENT and GL_PACK_ALIGNMENT to 1 once, right after the library's own EGL context is made current in init_opengl_ctxt(). Every texture upload and read back in the library goes through that context, so all formats now use packed rows and no call site has to remember to set it; the per-call toggles for 8bppYCbCr400 are removed. Formats whose rows are always a multiple of 4 bytes are unaffected.

Test: OpenGLContextTest.TextureUploadOfTightlyPackedRows uploads 400, 420, 422, 444 and RGB888 buffers at widths 1 to 10, reads the textures back and checks they hold exactly the source bytes. Built with -DUHDR_ENABLE_GLES=1 and AddressSanitizer on Mesa llvmpipe, the original code reports a heap-buffer-overflow in create_texture() on this test; with the change it passes, along with the rest of the unit tests.

Side note, not changed here: under GLES the existing EditorHelperTest cases for 12bppYCbCr420 pass planes[0] of an image whose Y, U and V planes are separate allocations to create_texture(), which expects the three planes in one buffer, so those cases read past the Y plane with or without this change.

[gemini-code-assist]

Code Review

This pull request configures the OpenGL context globally to use 1-byte row alignment for unpacking and packing during context initialization, and removes local alignment toggles in create_texture and read_texture. It also adds a new unit test TextureUploadOfTightlyPackedRows to verify texture upload and readback of tightly packed rows for various formats and widths. I have no feedback to provide as there are no review comments.