pkg/highway: replace resolveTargetQuery with direct CPU detection

The previous runtime_detect.zig called std.zig.system.resolveTargetQuery
which pulled in the entire Zig target/CPU model table infrastructure for
every architecture (~4,000 symbols, ~175 KB of data tables, ~130 KB of
code). This bloated the binary by ~500 KB and shifted code layout enough
to cause a measurable icache/branch-predictor regression in unrelated
hot paths like the terminal parser (~20% more cycles for identical
instruction counts).

Replace with minimal, direct CPU feature detection per architecture:
CPUID + XGETBV inline assembly on x86, sysctlbyname on Darwin AArch64,
and getauxval/prctl via std.os.linux (direct syscalls, no libc) on
Linux for AArch64, PPC, S390x, RISC-V, and LoongArch.

Split into per-architecture files under src/detect/ for
maintainability.

pkg/highway/src/detect/linux.zig

@@ -0,0 +1,10 @@
+/// Reads from the ELF auxiliary vector (set by the kernel at process
+/// start). Does not call into libc.
+pub inline fn getauxval(key: usize) usize {
+    return @import("std").os.linux.getauxval(key);
+}
+
+/// Direct syscall wrapper for prctl(2).
+pub inline fn prctl(option: i32, a2: usize, a3: usize, a4: usize, a5: usize) usize {
+    return @import("std").os.linux.prctl(option, a2, a3, a4, a5);
+}