Handle buffer allocation failures

[vinioliveiras]

Handle failed GPU buffer allocations instead of crashing

Refs #3018.

When the GPU runs out of memory, Wayfire currently crashes: several places
ignore buffer_reallocation_result_t::FAILED (or a NULL texture / render
pass from wlroots) and render into nothing. This is easy to hit on NVIDIA,
whose Linux driver does not spill VRAM into system memory: a game using
~6.5 of 8 GB was enough to take the whole session down within minutes.

Following the suggestion in #3018 (disable the effect when there isn't
enough memory), this skips the effect or the frame instead:

core

• render_pass_t::clear/add_rect/add_texture: no-op without a pass
  (wlr_renderer_begin_buffer_pass() returned NULL, "Failed to create FBO")
  or without a wlr_texture.
• texture_t::get_width/height, gles_texture_t: tolerate a NULL texture
  (texture id 0, draws nothing).
• transformer_base_node_t::get_updated_contents() returns nullptr when
  the buffer or its texture can't be made; 2D/3D transformers skip the frame.
• View snapshots (close animation) are skipped when the buffer fails.

plugins

• blur: no saved pixels / blur passes → the window is drawn without blur.
• animate (unmapped-view snapshot, squeezimize), workspace-wall, grid
  crossfade: skip.

Tested on wlroots 0.20.2 with an RTX 4060 Laptop (8 GB): a game kept the card
at 7.8–7.9 GB for long sessions; the log showed thousands of
Failed to allocate auxilliary buffer / gbm_bo_create failed lines and no
crash, where before the session died (backtraces below). Some crashes this
fixes:

Fatal error: Segmentation fault
#3  wlr_render_pass_add_rect
#4  wf::render_pass_t::clear(wf::regionf_t const&, wf::color_t const&)
#5  wf::render_pass_t::run_partial()
#6  wf::render_pass_t::run(wf::render_pass_params_t const&)
#7  wf::scene::transformer_base_node_t::get_updated_contents(...)
#8  wf::scene::blur_render_instance_t::render(...)::{lambda()#1}::operator()() const

(preceded by [render/gles2/renderer.c:114] Failed to create FBO)

and Cannot run a render pass without a valid target aborts from
view_interface_t::take_snapshot() during the close animation.

Happy to split this further or change the approach.

[ammen99]

I wonder whether it won't be simpler to preallocate a small 1x1 buffer and return that in case of failed allocations. Visually it won't work, asviews with transformers will be reduced to 1x1 buffers so likely will look broken. On the upside, plugins don't have to handle the failure case explicitly.

The alternative would be a bigger refactor of most plugins. For example, a view transformer might detect that an allocation has failed, and if so, then defer rendering to its child nodes directly and skip applying the transformation. The effect is not visible, but at least the view does not disappear. For plugins like Cube and Expo, we'd have to pre-allocate all buffers we need and not start the effect if they are not allocated.

Maybe for now we can try the first approach - it will 'fix' crashes in all plugins. Then, we can update plugins one by one to handle this in the 'correct' way.