Revamp MSAA images and render passes.

[floooh]

This brings the API closer to Metal and WebGPU and addresses a couple of long-standing issues:

• remove the hidden 'msaa vs resolve' magic in sg_image and sg_end_pass()
  • the msaa-resolve texture is no longer implicit part of an sg_image with sample_count > 1, instead the msaa and resolve textures are now separate sg_image objects
  • sg_pass_attachment_desc gets a separate resolve image entry, if this is present, an MSAA resolve will be performed in sg_end_pass() => NOTE: depth/stencil attachments cannot be MSAA-resolved (see WebGPU spec), for color attachment resolve, there will be a separate sg_pass_desc.resolve_attachments[] array.
• sg_action is split into sg_load_action and sg_store_action, where sg_store_action only contains dontcare and store (this is similar to WebGPU, but not similar to Metal - e.g. the Metal store actions 'resolve' and 'resolve-and-store' are implicit by the presence or absence of a resolve attachment
• depth/stencil image objects should now be regular textures in the GL backend (if this means dropping WebGL support, so be it)

...optional: get rid of baked pass objects completely? It's unclear whether would actually still be needed even with a Vulkan backend in place. => still a good idea in the GL backends, because framebuffer setup is pretty elaborate.

Related issues and PRs:

• [Reminder]: expose Metal's render pass store action in public API? #563
• _sg_mtl_begin_pass: explicit store action for depth and stencil attachments (MTLStoreActionStore) #815
• Reading from depth textures #576
• Depth SG_IMAGETYPE_CUBE in Metal #545
• sokol_gfx.h: investigate depth texture support #165
• Experimental depth read support #591

Also see:

• Sampling from MSAA textures gpuweb/gpuweb#990

[floooh]

UPDATE: Fixed in the recent master branch commits, the NSOpenGLContext generally wasn't active inside event handlers.

The new _sg_gl_create_image() now crashes in the mrt-sapp sample when resizing the window because glGenTexture() fails to create a texture, wtf?

=> more things wrong, in the webgl2 version the sokol-gfx validation triggers for trying to bind an MSAA image as texture (this was an actual bug)

...this seems to be a separate issue, also happens on the master branch... need to bisect whether this is a sokol-gfx GL backend regression, or a macOS problem (the SAPP_EVENTTYPE_RESIZED event is called from within the NSOpenGLView.reshape method, maybe there's not GL context bound in this function => the glGenTextures function which fails is the first GL function in that code path).

[floooh]

PR: #834