VIDEO/D3D12: GPU tonemap path for HDR screenshots

Counterpart to 5fbe019 (D3D11 GPU tonemap).  Moves the HDR->SDR
conversion in d3d12_gfx_read_viewport from the CPU decoder in
dxgi_hdr_readback_to_bgr24() to a GPU render using the shared
PSMainToSDR entry point in hdr_sm5.hlsl.h.

The GPU path:
 - Lazily creates a dedicated tonemap PSO into d3d12->hdr.pso_readback
   on the first HDR screenshot.  The PSO targets B8G8R8A8_UNORM, reuses
   the existing HDR root signature and vertex layout, and runs the
   PSMainToSDR pixel shader.
 - Allocates a transient HDR source texture (PIXEL_SHADER_RESOURCE) and
   copies the swapchain backbuffer into it; the swapchain itself has no
   SRV bind flag.
 - Allocates a transient SDR render target using a fixed RTV slot
   reserved at heap init time (one past the last per-pass slot).
 - Records a command list that does the tonemap draw, transitions the
   SDR RT to COPY_SOURCE, and copies into a D3D12_HEAP_TYPE_READBACK
   buffer with a GetCopyableFootprints-computed layout.
 - Signals the queue fence, waits, maps the readback buffer, and
   unswizzles BGRA8 -> BGR24 bottom-up.

On any failure (PSO compile, OOM, resource creation, Map) the caller
logs RARCH_WARN and falls through to the existing CPU decoder.  HDR
screenshots keep working even if the GPU path breaks on a particular
driver or hardware combination.

RTV heap size is bumped by 1 to reserve the readback slot.  All other
HDR-era screenshot infrastructure (UBO, vertex buffer, root signature,
sampler) is reused as-is.

Author: LibretroAdmin