fix: repair four pre-existing main CI failures (#3135)

[localai-org-maint-bot]

fix: repair four pre-existing main CI failures (#3135)

Four independent failures on main block all open pull requests.

macos-metal-mlx: SpeechEngine explicitly defaulted its copy and move
special member functions, but a std::mutex member makes them implicitly
deleted. Clang promoted -Wdefaulted-function-deleted to an error under
-Werror. Changed all four to = delete.

sanitize-cpu address: the elementwise GEMM scalar fallback dereferenced
unaligned uint16_t pointers in ElemA::Cvt and F16ToF32 at 11 call
sites across four files. Wrapped each dereference with
vt::LoadUnaligned for a safe scalar load.

sanitize-cpu thread: the gemma4_moe FP8-native branch at line 1637
matched on CPU, called EnsureGemma4Fp8NativeOnDevice which returns false
on every non-ROCm build, then had no fallback. The output buffer y was
never written. Under TSan the allocator zero-fills memory, exposing the
use of uninitialized memory as all-zeros output. Added an else clause
that falls back to the same host computation path the adjacent branch
uses.

windows-msvc: host_available_memory_bytes passed the glibc-only "e"
(close-on-exec) flag to fopen. MSVC UCRT treats unknown mode flags as an
invalid parameter and calls __fastfail, killing the process with
0xC0000409 before fopen returns. Changed the mode to "r" so fopen returns
nullptr on Windows where /proc/meminfo does not exist, and the existing
nullptr guard returns 0 as callers expect. The fopen fix exposed a
latent second issue: the post-build CRT audit calls dumpbin directly, but
dumpbin is not on PATH on windows-2022 runners (the VS cmake generator
finds cl.exe internally without exporting the toolchain). Added a
vswhere lookup that prepends the MSVC bin directory to PATH before the
CRT audit runs. The vswhere lookup alone was not sufficient: dumpbin calls
used 2>&1 to merge stderr into stdout, but under $ErrorActionPreference =
"Stop" (line 13) PowerShell 7 wraps native command stderr as ErrorRecord
objects and promotes them to terminating errors, swallowing dumpbin's
stdout and leaving the directive output empty. Removed 2>&1 from all
dumpbin and cl call sites, added /nologo to suppress the banner, and
added a dumpbin availability check after the vswhere block. The cl banner
call temporarily relaxes ErrorActionPreference to Continue since cl writes
its banner to stderr.

Closes #3135

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:regolo-glm5.2 [maki]

[localai-org-maint-bot]

Closing as redundant. The four CI failures this PR addressed have all been fixed independently:

• macos-metal-mlx (SpeechEngine deleted members) — fixed in fix(CI-SPEECH-ENGINE-DELETE): match code to comment on deleted special members #3136, merged
• sanitize-cpu address (unaligned GEMM scalar tail) — fixed in fix(KERNEL-GEMM-CPU-ELEM-X86WIDE): replace typed scalar-tail dereferences with LoadUnaligned #3138, merged
• sanitize-cpu thread (gemma4_moe else-block) — fixed in fix(ENG-EXPERT-STREAM): fall back to host BF16 when FP8-native device arm refuses #3141, merged
• windows-msvc-cpu — confirmed not a required check (feat(BACKEND-ROCM-QUANT-GATHER): gather packed embeddings on ROCm #3097 merged while it was failing)

The enforcement floor advance (#3142) is tracked separately and supersedes any remaining main-CI repair work.