[Core:Bugfix] Fix crashes when mmap allocation fails across all backends

[v0jiuqi]

When mmap-backed allocation fails (disk full, mmap error, file too short), backends silently stored null device pointers into tensors, causing EXC_BAD_ACCESS / SIGBUS crashes on iOS.

Fixes across all backends:

• LLM: grow mmap cache file before mmap to avoid SIGBUS past EOF; release module before runtime in load() to fix use-after-free
• Metal: check nil Metal buffer from failed mmap, fall back to non-mmap GPU allocator (mStaticAllocatorRaw)
• OpenCL: null-check all alloc() paths in onAcquire, fall back to non-mmap BufferPool; propagate MmapPool failure instead of wrapping null
• CPU: fall back to mStaticAllocatorRaw when mmap static allocation fails
• Vulkan: fix CALL_VK no-op in release builds; null-check allocations in onAcquire, VulkanBuffer/VulkanImage constructors
• CUDA: add null checks after pool->alloc() in all 26 execution kernels, return OUT_OF_MEMORY or set mValid=false

Discussed-in: Merge-Request 28543428 , URL: https://code.alibaba-inc.com/AliNN/AliNNPrivate/codereview/28543428
GitOrigin-RevId: b59217536d041370f22167289464ad7225a14178

Description

Module

Type

• Feature
• Bugfix
• Perf
• Refact
• Style
• Doc
• Test
• Chore

Checklist

• Commit message follows [Module:Type] Description format
• Code compiles without errors
• Tested on relevant platform(s)
• No unrelated format or style changes included

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

jingbang.yjb seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}