[d3d9] Bind framebuffer rework

[K0bin]

Successor to the scary half of #5092

Fixes #5116

First of all it rewrites the D3D9DeviceEx::BindFramebuffer logic based on what I found during testing:
Ignoring binds based on the write mask is an exception for 1x1 RTs.
If there's a size mismatch across render targets (and/or with the DS) the draw just gets skipped on Windows.
I didn't implement the skipping because it's another thing that would be checked on every draw and not a single game has hit that case so far. So I'll just ignore that until it becomes necessary.

After that there's another optimization to avoid needless layout transitions.
In the last couple of months, I optimized the framebuffer binding logic to bind
the render targets that the game tells us to bind instead of trying to be smart
and only binding the ones that the shader writes to and that have a non-zero write mask.
The goal behind this was to avoid having to break up render passes based on state or
shader changes. Besides that, we don't use VK_IMAGE_LAYOUT_DEPTH_READONLY_STENCIL_ATTACHMENT_OPTIMAL
anymore for the render pass depth stencil attachment. It does get used as the default layout for depth stencil images that can be sampled.
Those changes massively reduced the amount of render passes (and barriers) in some games.
The exception to either of those two is that the same texture is bound for sampling and as an RT or DS.
In that case we first try to unbind it in the case of RTs or bind it as DEPTH_READONLY in the case of DS
if possible (shader write masks, blend state write masks/depth write state allow it) before pulling out
the big guns with LAYOUT_FEEDBACK_LOOP.

Dead Space 2 has a deferred lighting pass where it ping pongs between two shaders.
One shader is very simple and only used to update the stencil buffer.
The other shader samples the depth buffer and renders something to a render target.
Neither of those writes to the depth buffer.
Both of those use the same depth stencil surface but because of the different shader
sampler masks, we transition the depth buffer to READONLY when the shader, that samples it,
gets bound. Then when the next shader gets bound, the "hazard" is cleared because the sampler mask
indicates it's not sampled anymore. This makes us rebind the depth buffer as RW because the assumption is that binding as RW by default will allow us to avoid splitting the render pass for every change to the depth write state.
The game does ping-pong between those two shaders after a single draw call each, so we also switch between RW and READONLY after a single draw call each.

The PR solves this by only checking, whether the shader actually uses the texture bound at a slot, if there wasn't a hazard before or if the hazard cannot be resolved by unbinding the RT (or binding the DS as readonly).
This reduces the number of render passes at the beginning of Dead Space 2 from 70 to 14 and later parts of the game see similar improvements.

[K0bin]

Briefly tested with:

• Risen
• Dead Space 1 (with the 1x1 shadows mode)
• Dead Space 1 (with the NULL format shadows mode)
• Dead Space 2
• Assassins Creed 2
• Fallout New Vegas
• GTA IV
• Half Life 2
• Mirrors Edge
• Remember Me
• Silent Hill 2 (with Enhanced edition mod)
• Skyrim
• Star Craft 2
• Star Wars Republic Commando
• Team Fortress 2
• Tomb Raider Legend
• Track Mania Nations Forever
• Witcher 2

[misyltoad]

Looks fine in theory. Definitely needs more stress testing though