perf(agc): keep the cached address space for a read-only V# across two imports - #5428

Two things before this can go in:

• Numbers for this change alone: the summary lines of main and of main plus this PR from the same scene, with the traces.
• The copied read-only region now overlaps the space's live ranges. A writable V# that lies inside one of those ranges doesn't dissolve the space, so it binds the import while the read-only V# binds a copy of the same bytes (addDescriptorRegion, GuestBufferMemory.cpp:2477). On main both use the one merged region. Please keep the dissolve when a write range overlaps the copied region, or add a test that shows the two views stay coherent.

Rebased onto main to resolve the conflict in tests/BdaResources.cpp after #290; no other change.

Other open pull requests touch the same code. Merging one will break the other, or both implement the same thing. If this one builds on them, list them in Depends on.

@oneandonlydean thanks for e5d6556 and fd2c300, the dissolve for a written V# over the copied read-only range and its test. That covers the coherence point.

On the numbers: I can't take a main capture of the Kena (PPSA01802) scene yet. Plain main (aa19278, this PR's base) aborts before the title scene, before the trace window opens:

[videoout] presentation failed: AGC graphics: shader buffer exceeds descriptor range limit
terminate called after throwing an instance of 'std::runtime_error'
  what():  AGC graphics: shader buffer exceeds descriptor range limit

It stops the same way with #5573 on top. My earlier numbers came from a local stack that skips draws which throw, which is why the description doesn't present them as main against this PR. I'm looking into the range-limit stop now. Once a base reaches the scene, I'll post the summary lines and traces of base and base plus this PR from the same session.

Pushed two commits on top so this can go in. If a written V# overlaps a read-only V# that is copied across two imports, the space now dissolves, in either order of the two, as it does on main without this branch. That way both bind the one merged region. If no written range overlaps, the space stays as before. The new spaceReadAcrossImportsWrittenTests in tests/BdaResources.cpp checks that the written and the read-only V# map the shared bytes through one buffer at matching offsets, and that the space dissolves. It fails on the previous head ("maps other bytes than it") and passes now. Full build is green on fd2c300.

One run of main + this PR, as an indication only (n=1, no main run from the same session to compare against): PPSA21564 first level, APS5_TRACE_FRAMES=4300:80, RTX 4090, NVIDIA 610.57.04, Linux, 3840x2160. Frame time p50 21.15 ms, p90 23.30 ms; GPU busy 14.7 ms per present. Gate phases: title 23.8, menu 20.3, level 27.0 fps, in line with recent main runs on this machine. The checkpoints look the same as on main.

This closes our earlier ask.

Resident Evil Village (PPSA01556) stops on this change's path at the title:

[videoout] presentation failed: AGC graphics: guest buffer view 0x1f0494592c+0x1b2190 exceeds its GPU owner 0x1f04940000-0x1f04950000 (buffer 0, direct 1, mirror 0)

Since this PR, a read-only V# that partly overlaps the cached space's base gets a region of its own, so a build's own regions can overlap base regions. GuestBufferMemory::owner() still assumed they don't. For this 1.7 MB view it returned the base region of one 64 KiB mapping, and Descriptor refused the view. A dissolve after UploadPrepare reaches the same lookup during the first scene load. #5815 fixes both. Its description has the details, a test that fails on main with this message, and the runs.

Setup for the stop: main 23be5a0 with #5688 and #5703, plus local test commits listed in #5815. Windows 11, RTX 3070 Laptop, driver 617.42.