perf(agc): release a heap mirror once a host import serves its range

[Celegans12]

What

An address-based build maps each readable registered range through a host import, else through a heap mirror, a host-visible copy kept current by write tracking. A range mirrored while APS5_HOST_IMPORT_MIB was full kept its mirror after a later pass imported it (budget freed by an unmap). Until the guest freed the range, the mirror held host-visible memory and bytes no build read.

• AcquireRegistered now releases the heap mirror of each range its pass serves through an import. Recorded work that bound it keeps it until it completes, as for a swept mirror. Image mirrors are unchanged.
• The [buffers] image mirrors line (APS5_PROFILE_DRAW) counts the released mirrors.
• In PPSA26344 v1.008 (main with our open PRs and feat(agc): host import windows, least-recent eviction under the budget #290), heap mirrors held 10.6 GiB in 37k ranges while host imports swung between 0.7 and 3.7 GiB with the mirror count flat.
• feat(agc): host import windows, least-recent eviction under the budget #290 rewrites importAllocation and its call in this pass (it conflicts with main already; this change adds no conflict). Its eviction drops imports of ranges that stay registered; the next pass then mirrors such a range anew from the guest bytes. fix(agc): throw instead of aborting when a heap mirror cannot be made #1085 (ours, acquireMirror and the same tests), fix(agc): release the current address space with the image mirrors #1071 (ClearImageMirrors) and fix(agc): retry failed host imports after range replacement #1357, fix(agc): order direct reads after copied writes #1359 and perf(agc): reuse import decisions across mapping changes #1361 (the import entry in the same AcquireRegistered loop) merge cleanly.

Tested

• Windows 11, MinGW-w64 GCC 15.2.0, based on main df16c4c: configure -DBUILD_TESTING=ON, full build, --target libs, ctest:
• 404/404 tests pass (ctest --output-on-failure)
• Same build on lavapipe (Mesa 26.2.4 llvmpipe, 8-wide subgroups, VK_DRIVER_FILES): 403/403 pass; agc_driver_flat_store is left out there because it takes several minutes on Windows llvmpipe
• agc_driver_graphics (mock device, new case next to the heap mirror tests): a mirrored range that a later build imports loses its mirror, an imported range is not mirrored again, and a build without the import mirrors the current guest bytes anew. Fails on main.
• PPSA26344 v1.008 (with our open PRs, 3,000 s): 697 heap mirrors (862 MiB) released for imports, all in the first 923 s; heap mirrors then stayed under 5.7 GiB and no vkAllocateMemory failed.
• Not tested: Linux (CI), console behaviour.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed in TechnicalDebt
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• AI-assisted: yes

[EdouardSence]

Tested on AMD under Linux, since this PR wasn't: RX 9070 XT (RDNA4, GFX1201), RADV / Mesa 26.2.4, Linux 7.2.9 (CachyOS), GCC Release build.

• ctest: this PR merged onto main 07c2405, all 452 tests pass with ctest -j8 (main alone: 452/452 on the same machine).
• Astro Bot (PPSA21567, retail): built as main + feat(agc): implement deferred faults for reads through null SRT pointers #759 (needed to boot this title on main) + this PR, then one scripted run headless from boot through the title screen. No error and no amdgpu fault came from this change.
  • Presents per 10 s on the title screen: 243, 231, 227; with main + feat(agc): implement deferred faults for reads through null SRT pointers #759 alone: 226, 229, 226.
  • The run then stops, with and without this PR, on storage image access to depth/stencil surface ... is not implemented (the gap feat(agc): depth planes sampled as 2D arrays, cubes and integer views #425 fills). So nothing past the title screen is covered.

AI-assisted: the runs and this write-up were done by Claude (Claude Code) on my machine.