linalg/x86_64: build on the 1.91 MSRV and the cfg-off path

The AMX / AVX-VNNI detection reads CPUID via std::arch::x86_64::__cpuid_count.
That intrinsic is `unsafe` on tract's MSRV (rustc 1.91) but was made safe in a
later release, so the calls compiled locally yet broke every 1.91 CI job with
E0133. Wrap each call in `unsafe { }` (required on 1.91) and add
`#[allow(unused_unsafe)]` so newer toolchains, where the call is safe, don't
trip `unused_unsafe` under `-D warnings`.

Also gate the `PackedI8K4` and `super::amx` imports in mmm.rs on the cfgs that
actually use them, so the old-assembler build (Debian stretch, all kernel cfgs
off) has no unused-import warnings.

Verified: tract-linalg compiles on rustc 1.91.0 and 1.94, clippy-clean.

https://claude.ai/code/session_018Hes6yEvk2TSWB26SAJfqT

Author: czoli1976

linalg/src/x86_64_fma/amx.rs

@@ -52,7 +52,8 @@ pub fn cache_sizes() -> CacheSizes {
     *CACHE.get_or_init(|| {
         let mut out = CacheSizes::default();
         for sub in 0..16 {
-            let r = std::arch::x86_64::__cpuid_count(4, sub);
+            #[allow(unused_unsafe)]
+            let r = unsafe { std::arch::x86_64::__cpuid_count(4, sub) };
             let cache_type = r.eax & 0x1F;
             if cache_type == 0 {
                 break;
@@ -83,7 +84,8 @@ fn cpu_has_amx_int8() -> bool {
     if !std::is_x86_feature_detected!("avx512f") {
         return false;
     }
-    let r = std::arch::x86_64::__cpuid_count(7, 0);
+    #[allow(unused_unsafe)]
+    let r = unsafe { std::arch::x86_64::__cpuid_count(7, 0) };
     // bit 24 = AMX-TILE, bit 25 = AMX-INT8 in EDX.
     const AMX_TILE: u32 = 1 << 24;
     const AMX_INT8: u32 = 1 << 25;

linalg/src/x86_64_fma/amx_bf16.rs

@@ -30,7 +30,8 @@ fn cpu_has_amx_bf16() -> bool {
     if !std::is_x86_feature_detected!("avx512f") {
         return false;
     }
-    let r = std::arch::x86_64::__cpuid_count(7, 0);
+    #[allow(unused_unsafe)]
+    let r = unsafe { std::arch::x86_64::__cpuid_count(7, 0) };
     const AMX_BF16: u32 = 1 << 22;
     const AMX_TILE: u32 = 1 << 24;
     (r.edx & AMX_BF16) != 0 && (r.edx & AMX_TILE) != 0

linalg/src/x86_64_fma/avxvnni.rs

@@ -26,11 +26,13 @@ fn cpu_has_avxvnni() -> bool {
     if !std::is_x86_feature_detected!("avx2") {
         return false;
     }
-    let max_sub = std::arch::x86_64::__cpuid_count(7, 0).eax;
+    #[allow(unused_unsafe)]
+    let max_sub = unsafe { std::arch::x86_64::__cpuid_count(7, 0) }.eax;
     if max_sub < 1 {
         return false;
     }
-    let r = std::arch::x86_64::__cpuid_count(7, 1);
+    #[allow(unused_unsafe)]
+    let r = unsafe { std::arch::x86_64::__cpuid_count(7, 1) };
     (r.eax & (1 << 4)) != 0
 }
 

linalg/src/x86_64_fma/mmm.rs

@@ -2,8 +2,11 @@ use crate::Ops;
 use crate::block_quant::*;
 use crate::mmm::ImplementationQuality::ManuallyOptimized;
 use crate::mmm::MatMatMul;
-use crate::pack::{PackedFormat, PackedI8K4};
+use crate::pack::PackedFormat;
+#[cfg(any(tract_avx512vnni, tract_avxvnni, tract_amx_int8))]
+use crate::pack::PackedI8K4;
 
+#[cfg(tract_amx_int8)]
 use super::amx::{PackedAmxA, has_amx_int8};
 #[cfg(tract_amx_bf16)]
 use super::amx_bf16::{PackedAmxBf16A, PackedBf16K2, has_amx_bf16};