GLTextureUploadMeta process could be optimized in a way where the associated GstVaapiTexture to the native texture supplied by the user through _upload() could be cached.

Currently, even if the GLTextureUploadMeta is attached to the buffer during its creation in the parent pool, the _upload() function would check for an existing GstVaapiTexture locally only, i.e. in that specific meta instead of its active neighbours too (in the pool).

This means that, say that GLTextureUploadMeta TUM1 and TUM2 are respectively attached to buffers B1 and B2, if B1 is uploaded with Texture id Tex0, and so is B2 afterwards with the very same Texture id Tex0, then we are still have two separate GstVaapiTexture holders for the same native Tex0. This is extra memory storage we don't really want.

Imagine we create a pool of 8 buffers, that end up with underlying resources (e.g. FBO, pixmap) requirements for 1080p (thus, x8), this starts to add up for the very same Tex0.