v0.3.0

A performance release. Everything below was measured on one machine (Core i5-14600KF, RTX 3080, Windows 11), with
runs taken in turn (A, B, A, B...) because the same frame varies 10-18 ms of GPU time between runs. Not measured on
other graphics cards (AMD, Intel, integrated): feedback is welcome in an issue.

• The graphics work runs on its own thread (on by default). The game's calls into the Vulkan backend used to run
  on the game's own thread, in the chain that also runs its logic. Unlimited frame rate on the Pewter City route:
  Eevee 126.9 -> 185.5 fps (frame time ratio 1.46, 95% interval 1.40-1.53), Pikachu 132.1 -> 191.4 fps (1.45, 1.31-1.54);
  the 99th percentile frame time drops from 13.3 to 9.9 ms. At the game's own 30 fps cap you will not see more frames:
  the gain is CPU headroom. The paced 60 fps mode is
  unchanged (16.63 ms per frame with and without). Rendered frames match the old path within the run-to-run noise;
  the Vulkan validation layers report no errors on the regression routes of both games. render_thread = 0 in
  larecomp.ini (or LARECOMP_RENDER_THREAD=0) turns it off.
• Models are read from video memory on dedicated GPUs. The vertex and index data was read by the GPU in place from
  the game's memory (system memory, over PCIe) for every vertex invocation. It is copied to VRAM now: GPU time of the
  Pewter City frame 10.2 -> 1.2 ms. Integrated GPUs keep the old path (not tested). vertex_read = direct in larecomp.ini (or
  LARECOMP_VERTEX_READ=direct) goes back to it.
• Mutex waits no longer stall the game threads. A contended lock spun briefly before sleeping on a condition
  variable (~1000 contended locks per frame per thread, ~12 us each on Windows). The paced 60 fps mode now holds 60.0
  fps (it was 58.8-59.3); unlimited frame rate went from about 60 to 65-74 fps at the time.
• larecomp.ini edited by hand on Windows works: lines ending in CR+LF kept the CR in the value, so
  resolution = dynamic, upscaler = bilinear and the percentage modes fell back silently.
• Dynamic resolution (resolution = dynamic, ini only) judged a steady 30 fps as too slow and dropped to 0.60
  scale; it budgets by the scene's pace now.
• For developers: a much faster regression battery (6 routes of both games in ~100 s, was ~30 min), an interleaved
  A/B performance test with a confidence interval, per-draw GPU timing and a frame time CSV (docs/DEPURACAO.md,
  docs/PESQUISA-GPU.md).
• Upgrading from 0.2.x: unzip over your install; no new setup (the game library is not rebuilt).
• Tested with the zip built for this release, unzipped over 0.2.x installs of both games and started without MSYS2 on the PATH (no recompiling needed): Eevee loads a save and reaches the world (Route 2); Pikachu boots with the render thread on and plays the opening movie.

Updating: with the game closed, unzip over your folder (keeps the game files, saves, translations and settings).