Auris v1.2.1

fix(linux32): correct x87 FP precision and BMI2 guard for i386 builds

On Linux 32-bit, two independent issues caused garbage (!!!!) output:

1. GCC defaults to x87 (80-bit extended precision) for scalar floats
   on i386 even with -mavx2. Whisper's softmax/logit values diverge
   from every other platform. Fixed by adding -mfpmath=sse -msse2 to
   the Linux x86 build flags in premake5.lua.

2. _pdep_u64 is an x86_64-only intrinsic. The #ifdef __BMI2__ guard
   in quants.c was insufficient — on i386 with BMI2 the intrinsic is
   absent, producing silent garbage. Fixed by guarding with
   defined(__BMI2__) && defined(__x86_64__) so 32-bit hits the
   correct scalar fallback.

Patch applied to vendor submodule, CI linux32 jobs, and RUN_ONCE
setup scripts.