rend: optimize ARM64 framebuffer paths

[rtomasa]

Summary

• Add ARM64 NEON paths for framebuffer conversion and VRAM writes.
• Preserve scalar fallbacks for unsupported or unaligned regions.
• Validate the physical VRAM allocation before vector reads, preventing out-of-bounds access with both 8 MB and 16 MB VRAM configurations.

[flyinghead]

You can get rid of pvr_mark_framebuffer_write() because writing the framebuffer to VRAM is only done when EmulateFramebuffer is enabled. And in this mode the framebuffer writes don't need to be tracked, since the framebuffer is displayed on every vblank no matter what.

I likely forgot to gate the write check in pvr_write32p() behind the Internal template argument. I have some changes coming in this area so I'll fix it.

Everything else looks good to me although I haven't tested it.

[rtomasa]

Thanks for the feedback. I'm excited to test it out with all your corrections to my code.

[flyinghead]

I fixed pvr_write32p code.
Can you do the pvr_mark_framebuffer_write() change? Then I'll merge it

[rtomasa]

Done!