AGC: native 4-sample D32 depth sampled with HTILE but without ITERATE_256 is refused (Onimusha: Way of the Sword, PPSA27836) #6362

Onimusha: Way of the Sword (PPSA27836) samples a native 4-sample depth target whose HTILE isn't ITERATE_256, and stops on the new game's first scene load. No open PR covers this case. #6300, #6297, #5554 and #5627 all assume ITERATE_256.

AGC graphics: native multisample depth initial import requires pipe-aligned ITERATE_256 HTILE without VRS

I added a local print of the target in prepareDepthSampling (DepthSurface.cpp):

So the target passes every other check before this one: format, sample count, Z_X single slice, XOR, base alignment, no stencil. Only ITERATE_256 is off.

Where: after New Game, difficulty and control-style selection, during the first scene load. It happens every run.

Setup: Windows 11, RTX 5070 (NVIDIA 617.42), main 0afa6a6 with #5907, #5776, #5770, #5703, #5841, #5884, #5915 and #5815, plus local sample-count changes.

Host import: this refusal only shows with APS5_NO_HOST_IMPORT=1. With host imports on, the same load ends earlier, at the same point after the save, in VK_ERROR_DEVICE_LOST. NVIDIA device fault reports read invalid at 0x4840000, the same address every run. The last GPU checkpoint is an 8x8 compute program that samples two 2D textures and writes a 16x16x8 3D storage image. I'm still narrowing that one down. It isn't changed by any of these:

• APS5_NO_PIPELINE_LTO;
• reverting perf(agc): keep Windows host imports of write-watched private memory watched #6060;
• fix(agc): hold a guest allocation lease while a queue stores idle results #6246;
• APS5_TEXTURE_CACHE_MIB=6000;
• APS5_NO_DEFERRED_EVICTION.

How HTILE is laid out without ITERATE_256 isn't settled here, so I haven't guessed an implementation. If someone has the layout, or a hardware measurement of it, I can test a change on this title right away.