Narrow further, and don't use microarch

[jerinphilip]

This is some cursed debugging on GitHub CI trying to smoke out illegal instruction core dumped. Does not happen on my system, and ends up happening on CI. Happens intermittently. Did some guesswork attempts trying to improve the situation, did not pan out. So here we are. In the process I'm parallely discovering lldb, gdb and other weird places that I have no one but myself to blame for getting me into.

Hey, there's possibly light at the end of the tunnel.

* thread #1, queue = 'com.apple.main-thread', stop reason = EXC_BAD_INSTRUCTION (code=EXC_I386_INVOP, subcode=0x0)
    frame #0: 0x00000001002c4626 slimt-cli`slimt::Tensor slimt::qmm::detail::dot<(slimt::qmm::detail::Provider)3>(x=0x00007ff7bfefbcc0, W=0x0000621000007f88, a_quant=11.9759073, b_quant=72.2858505, name="rnn_o") at Gemmology.inl.cc:300:42
   297 	  float a_alpha = kInt8Maxf / a_quant;
   298 	  float b_alpha = kInt8Maxf / b_quant;
   299 	
-> 300 	  float bias_unquant_multiplier = (-1.0F * (a_alpha * b_alpha)) / kInt8Maxf;
    	                                         ^
   301 	  auto prepare_bias_callback =
   302 	      gemmology::callbacks::UnquantizeAndAddBiasAndWrite(
   303 	          bias_unquant_multiplier, bias.data<float>(),  //